The Bluetooth 6.3 module landscape is beginning to take shape following the Bluetooth Special Interest Group's (SIG) adoption of the Core 6.3 specification in May 2026. For B2B engineers designing products with wireless connectivity — whether in industrial automation, medical devices, asset tracking, smart building systems, or consumer-adjacent commercial hardware — understanding what Bluetooth 6.3 adds over the widely deployed Bluetooth 5.3 is a practical planning question, not just a spec-sheet exercise. This guide explains what each version is, what the four enhancements in Bluetooth 6.3 specifically deliver, and where the differences between 5.3 and 6.3 are genuinely consequential for B2B product design versus where they are incremental refinements that may not affect your application at all.

Summary

Bluetooth 6.3 builds on the Bluetooth 6.0 foundation introduced in 2024 with four targeted technical enhancements: more accurate and efficient Channel Sounding (the high-precision ranging feature), per-PHY Round-Trip Time accuracy declarations for better-optimised ranging, an expanded Host Controller Interface that future-proofs the architecture for new features, and harmonised RF requirements that simplify dual-mode radio design. For B2B products requiring precise indoor positioning, asset tracking, or access control, the Channel Sounding improvements in 6.3 are the most significant change. For products using dual-mode Bluetooth (both Classic and LE), the RF harmonisation reduces design complexity. Bluetooth 5.3 remains a mature, fully capable specification for the majority of B2B wireless connectivity applications where these specific capabilities are not required.

The Bluetooth version landscape: From 5.3 to 6.3

To understand where 6.3 sits, it helps to trace the recent version history briefly. Bluetooth 5.3, released in July 2021, is the current foundation for the majority of deployed Bluetooth LE modules. Its headline additions were Connection Subrating — which enables rapid duty-cycle switching on established LE connections to reduce idle-state power consumption — alongside improvements to Periodic Advertising efficiency, encryption key size management, and channel classification. Bluetooth 5.3 is a well-understood, widely supported specification and remains the right choice for many product designs.

Bluetooth 6.0, adopted in September 2024, introduced the major architectural additions of this generation: Channel Sounding (high-precision distance measurement), Decision-Based Advertising Filtering, Monitoring Advertisers, ISOAL enhancements for LE Audio latency, the LL Extended Feature Set, and a Frame Space Update. Bluetooth 6.0 and 6.1 are deployment upgrades that make Bluetooth smarter at ranging, scanning, privacy, and energy use, rather than chasing data rates.

Bluetooth 6.1 followed in May 2025 with privacy and power efficiency refinements. Bluetooth 6.3, adopted in May 2026, refines and extends the 6.0 foundation with four specific enhancements, all of which build on capabilities introduced at 6.0 rather than introducing entirely new feature categories. Bluetooth Core 6.3 introduces new features that boost ranging precision, expand interface capacity, and improve radio efficiency.

What Bluetooth 5.3 offers

Before examining what 6.3 adds, it is worth being clear about what 5.3 already provides well, since the choice of module version for a B2B product should be driven by application requirements rather than version numbers.

• Low energy connection management: Connection Subrating allows a device to maintain an LE ACL connection while rapidly switching between low-power idle and active duty cycles, delivering efficient power use across a wide range of peripheral and IoT device types
• Reliable short-range wireless connectivity: Bluetooth 5.3 supports data rates up to 2 Mbps (LE 2M PHY), ranges up to 400m in line-of-sight conditions (using LE Coded PHY), and simultaneous connections to multiple devices
• LE Audio readiness: Bluetooth 5.3 is fully compatible with the LE Audio specification, supporting LC3 audio codec, multi-stream audio, Auracast broadcast audio, and hearing aid profiles — a significant capability set for commercial audio and assistive technology applications
• Mature ecosystem: a large selection of certified modules from multiple manufacturers, well-established tooling, broad OS and stack support, and a substantial body of B2B deployment experience
• Basic direction finding: Angle of Arrival (AoA) and Angle of Departure (AoD) for approximate direction finding were introduced at Bluetooth 5.1 and are present in 5.3

For most short-range data connectivity applications — sensor nodes, HMI interfaces, industrial handheld devices, connected medical equipment, retail terminals — Bluetooth 5.3 is a fully sufficient specification today.

What Bluetooth 6.3 adds: The four enhancements

Bluetooth 6.3 does not introduce new connection types, higher data rates, or extended range. Its additions are targeted improvements to specific capabilities. Each is described below in terms relevant to B2B product engineers rather than at the full specification level.

1. Channel Sounding Inline PCT Transfer

Channel Sounding, introduced in Bluetooth 6.0, enables two devices to measure the physical distance between them with centimetre-level accuracy using Phase-Based Ranging (PBR) — a significant advance over the approximate RSSI-based proximity estimation that earlier Bluetooth versions supported.

The core challenge in Phase-Based Ranging is eliminating the phase offset introduced by each device's independent local oscillator (LO). In Bluetooth 6.0 and earlier Channel Sounding implementations, this is handled by a two-way digital cancellation process: the initiator and reflector each measure phase, exchange data, and the initiator combines both measurements digitally to cancel the LO error. This works, but introduces data overhead, post-processing latency, and sensitivity to LO drift during the exchange period.

Bluetooth Channel Sounding Inline PCT Transfer enhances Channel Sounding accuracy and efficiency by allowing the reflector to transfer phase-aligned tones directly into hardware, adjusting the phase of the tone at the reflector and eliminating excess result data, reducing overhead, and improving the speed and efficiency of the Channel Sounding procedure. In practical terms, LO offset cancellation moves from digital post-processing to analog hardware at the reflector, which reduces per-measurement data payload, lowers processing latency, and makes the distance estimate less sensitive to LO drift during the exchange. More Channel Sounding procedures can be completed in a given time window, and the distance result is available faster after the exchange completes.

For B2B product designers working on real-time location systems (RTLS), proximity-based access control, contactless tooling measurement, or any application where Bluetooth distance measurement accuracy and update rate matter, this is the most operationally significant change in 6.3.

2. Channel Sounding PHY-specific RTT Accuracy

Channel Sounding also supports Round-Trip Time (RTT) ranging — a complementary method that estimates distance by measuring the time of flight of an RF signal. Prior to 6.3, a device declared a single RTT accuracy requirement applicable across all PHYs, which did not account for the different radio characteristics of LE 1M and LE 2M PHYs.

Channel Sounding PHY-specific RTT Accuracy enables devices to declare Round-Trip-Time accuracy separately for each PHY rather than relying on a single value across all modes, so systems can select the optimal PHY-accuracy combination, improving precision, performance scaling, and interoperability in multi-PHY ranging scenarios.

For product engineers, this means Channel Sounding deployments on the faster LE 2M PHY can now be configured to perform the minimum number of exchanges needed for the required precision on that PHY, rather than defaulting to the more conservative LE 1M parameters. The result is lower radio-on time per ranging measurement, reduced power consumption, and better exploitation of LE 2M's capabilities in precision-ranging applications.

3. Running Out of Bits (HCI Expansion)

This enhancement is architectural rather than directly functional, but it has meaningful implications for long-term product platform decisions. The Host Controller Interface (HCI) is the standardised communication layer between a Bluetooth Host (typically the application processor and its software stack) and the Bluetooth Controller (the radio and link layer hardware). Commands and events between them are tracked using bitmask fields in the HCI.

The Running Out of Bits enhancement extends the architectural limits of the HCI by expanding the Supported Commands bitmask from 64 to 251 octets and the LE Event Mask from 8 to 255 octets, providing the necessary protocol capacity for future feature growth, ensuring discoverability of newly assigned commands and events, and maintaining backward compatibility through versioned commands and conditional support rules.

For B2B engineers, the practical implication is straightforward: a module or controller built to Bluetooth 6.3 has the HCI headroom to support all future Bluetooth SIG feature additions without requiring an architectural change to the interface itself. Products built on 5.3-era controllers may encounter HCI capacity constraints as Bluetooth continues to evolve. This is not an immediate concern for most current products, but it is a relevant consideration for platforms intended to remain in production and service for five to ten years.

4. ACP and C/I Limit Relaxation for Dual-Mode Radios

This enhancement addresses a design challenge specific to Bluetooth Dual-Mode devices — products that implement both Bluetooth Classic (BR/EDR, used for audio streaming, serial port profiles, and other Classic Bluetooth applications) and Bluetooth LE on the same radio.

Prior to 6.3, the Adjacent Channel Power (ACP) and Carrier-to-Interference (C/I) ratio requirements for Bluetooth BR/EDR were more stringent than those for Bluetooth LE, by around 15–17 dB in some offset ranges. This meant dual-mode radio designers had to meet the most conservative requirements of either standard, resulting in over-constrained RF designs that consumed more silicon area, required more complex transmitter and receiver architectures, and used more power than was technically necessary.

ACP and C/I Limit Relaxation harmonises RF requirements between Bluetooth Classic and Bluetooth LE by aligning BR/EDR Adjacent Channel Power and Carrier-to-Interference limits with the LE 1 MS/s framework, simplifying transmitter and receiver design targets and enabling more power-efficient, flexible RF architectures without compromising coexistence performance.

For B2B product engineers designing hardware that uses both Classic Bluetooth (for audio profiles, SPP, or high-throughput file transfer) and BLE (for low-power IoT connectivity, GAP/GATT services, or LE Audio), this simplifies the RF design problem and opens the door to more power-efficient dual-mode module designs from silicon vendors.

Bluetooth 5.3 vs Bluetooth 6.3: Key differences at a glance

Benefits of Bluetooth 6.3 modules for B2B products

• Centimetre-level indoor positioning: the Channel Sounding improvements in 6.3 make precision RTLS applications — where RSSI-based proximity is insufficient — more practical and accurate in real-world deployment conditions, directly opening use cases in asset management, personnel safety zones, and access control
• More efficient ranging for battery-powered devices: PHY-specific RTT accuracy reduces the number of ranging exchanges needed on capable PHYs, lowering radio-on time and extending battery life in position-aware wearables, tags, and IoT sensors
• Simplified dual-mode radio design: harmonised ACP and C/I limits reduce the RF engineering burden for products implementing both Classic Bluetooth and BLE, potentially enabling smaller, cheaper, and more power-efficient dual-mode chips in future module generations
• Architectural future-proofing: the expanded HCI removes a structural bottleneck that would otherwise have required module or controller replacements as new Bluetooth features are specified in future releases
• Competitive product differentiation: for B2B product designers serving markets where precision ranging is a differentiator — logistics, healthcare, retail, manufacturing — 6.3's Channel Sounding refinements are directly marketable capability improvements

Applications where Bluetooth 6.3 matters most

Real-time location systems and asset tracking

RTLS applications in hospitals, warehouses, and factories require sub-metre positioning accuracy that RSSI proximity cannot reliably provide. Bluetooth 6.3's Channel Sounding improvements — faster ranging, lower overhead, and reduced susceptibility to local oscillator errors — make Bluetooth a more competitive option in this market alongside UWB. For asset tracking deployments that already use Bluetooth infrastructure, 6.3 modules enable a meaningful accuracy upgrade without changing the wireless standard.

Access control and proximity-based security

Door access systems, secure zones, and equipment authorisation applications benefit from distance estimates that are resistant to relay attacks — a vulnerability of RSSI-based systems where the signal strength can be spoofed or amplified. Channel Sounding's time-of-flight approach provides a physically grounded distance measurement that is significantly harder to manipulate than RSSI, making 6.3 modules relevant for physical security applications that currently use NFC or RFID for close-range authorisation.

Industrial measurement and tooling

Contactless distance measurement, robotic arm positioning confirmation, and machine gap verification are applications where centimetre-level Bluetooth ranging, if achieved reliably, replaces more expensive or more complex measurement approaches. The reduced overhead of 6.3's Channel Sounding means more measurements per second in industrial automation contexts where update rate matters.

Dual-mode audio and industrial handheld devices

Products combining Classic Bluetooth audio streaming with BLE for configuration, OTA updates, or IoT connectivity — rugged handheld scanners, industrial headsets, ATEX-rated communication devices — benefit from the RF simplification that 6.3's ACP and C/I harmonisation enables. Simpler dual-mode radio architectures reduce design risk and component cost in this product category.

Long-lifecycle industrial and medical products

Products with five-to-ten-year design lives benefit from the HCI expansion even if they do not immediately use the new features. A 6.3-based module will support all future Bluetooth SIG feature additions without architectural rework, reducing the risk of a mid-lifecycle component obsolescence event driven by HCI capacity rather than genuine performance limitations.

Hardware availability: What to expect in 2026

It is important to be clear about the current state of the market. Bluetooth 6.3 was adopted by the Bluetooth SIG in May 2026, making it a very recent specification. Silicon vendors typically require six to twelve months from specification adoption to produce certified chips, and module manufacturers take additional time to design, certify, and bring to market modules based on that silicon. As of mid-2026, Bluetooth 6.3 certified hardware is beginning to emerge, with broader module availability expected through late 2026 and into 2027.

This matters practically for B2B engineers making module specifications decisions today. For products with short design cycles targeting a 2026 production date, Bluetooth 5.3 or 6.0 certified modules remain the pragmatic choice — the ecosystem is mature, tooling is established, and certification processes are well understood. For products with longer development cycles or scheduled design refresh windows in 2027, specifying 6.3 from the outset makes sense, particularly for applications where Channel Sounding or dual-mode RF improvements are in scope.

BCD Atlantik supplies a range of Bluetooth and Wi-Fi + Bluetooth combo modules from Feasycom, IOT747, Quectel, SparkLAN, and Flaircomm, spanning BT 5.0 through 5.2, with LE Audio capable modules and dual-mode audio modules available for current design programmes. As 6.3 certified modules become commercially available, these will be added to the range. The BCD Atlantik wireless connectivity range and our team can advise on the right module specification for your development timeline and application requirements.

Best practices for specifying a Bluetooth module for B2B products

1. Define your application's core wireless requirements first. Data rate, range, power budget, and whether precision ranging is required should drive module selection — not the latest version number. For most B2B connectivity applications, 5.3 is sufficient today.
2. Consider your product's design life and the upgrade window. For products in production through 2028 or beyond, 6.3's HCI expansion is a meaningful long-term advantage. For products shipping in 2026, validate that certified 6.3 modules are available on your development timeline before committing.
3. Evaluate Channel Sounding requirements honestly. Precision indoor ranging with Bluetooth is genuinely useful for a specific set of applications, but it requires both devices to support Channel Sounding and involves antenna and RF design considerations beyond the module itself. Assess whether your use case genuinely needs centimetre-level accuracy before making Channel Sounding a specification requirement.
4. Check certification requirements for your target market. Modules with pre-certified RF compliance reduce the regulatory burden, but this certification is version-specific. Changing the Bluetooth version mid-design may require re-qualification.
5. Engage your distributor early on module availability. For emerging specification versions like 6.3, lead times and availability windows can be longer than for mature variants. BCD Atlantik can advise on current availability, planned product introductions, and suitable bridge specifications for projects targeting Bluetooth 6.3 when development begins before full module availability.

Frequently asked questions

Is Bluetooth 6.3 backward compatible with Bluetooth 5.3 devices?

Yes. Bluetooth maintains backward compatibility across all versions. A Bluetooth 6.3 device will connect and communicate with Bluetooth 5.3, 5.x, and 4.x devices. However, the new features introduced at 6.3 — Channel Sounding improvements, PHY-specific RTT accuracy, and the expanded HCI — require both devices to support the relevant 6.x capabilities to be used. These features will not operate with a 5.3 peer device.

What is Bluetooth Channel Sounding and why does it matter for B2B applications?

Channel Sounding is a suite of ranging techniques introduced in Bluetooth 6.0 that lets two devices measure the physical distance between them with centimetre-level accuracy, using either Phase-Based Ranging or Round-Trip Time methods. This is significantly more precise than the RSSI-based proximity estimation available in earlier versions, and opens up use cases in asset tracking, access control, industrial measurement, and safety systems where approximate proximity is insufficient.

Does Bluetooth 6.3 improve data throughput or range compared to 5.3?

No. Bluetooth 6.3 does not change the maximum data rate (2 Mbps on LE 2M PHY) or the maximum range (up to 400m on LE Coded PHY) that were established in Bluetooth 5.0. Its enhancements are targeted at ranging precision, interface scalability, and dual-mode radio design — not raw throughput or link distance.

What does the HCI expansion in 6.3 mean for a product engineer?

The HCI is the interface between the Bluetooth radio controller and the host processor running the Bluetooth stack. The previous bitmask structure was approaching its capacity limit as Bluetooth added new features. The 6.3 expansion to 251-byte command and 255-byte event masks ensures that modules and controllers built to 6.3 can support all future Bluetooth SIG features without requiring a redesign of the Host-Controller interface. For products with long production lives, this reduces the risk of a forced hardware revision driven by HCI architecture constraints.

Are Bluetooth 6.3 certified modules available now?

Bluetooth 6.3 was adopted by the Bluetooth SIG in May 2026. Certified silicon and module availability is building through 2026, with broader commercial availability expected from late 2026 into 2027. For current design programmes, Bluetooth 5.3 and 6.0 certified modules remain the most accessible options. Contact BCD Atlantik to discuss availability and development timelines for 6.3-compliant modules.

Should I upgrade my existing 5.3-based product design to 6.3?

Only if a specific 6.3 capability addresses a requirement your current design cannot meet, or if a scheduled design refresh aligns with 6.3 module availability. Changing Bluetooth module specifications mid-programme involves recertification, potential re-qualification of RF compliance, and integration work. For products already in production or late-stage development, 5.3 remains fully capable for the vast majority of B2B wireless applications.

What is the difference between Bluetooth 6.0 and Bluetooth 6.3?

Bluetooth 6.0 introduced Channel Sounding, Decision-Based Advertising Filtering, Monitoring Advertisers, and several other new features. Bluetooth 6.3 refines two aspects of Channel Sounding (Inline PCT Transfer for efficiency, PHY-specific RTT accuracy for optimised ranging), expands the HCI to support future feature growth, and harmonises RF requirements for dual-mode radio design. Bluetooth 6.1 and 6.2 added intermediate refinements between these two versions. Bluetooth 6.3 is not a wholesale revision of 6.0 but a targeted set of improvements to specific capabilities.

Does BCD Atlantik supply LE Audio capable Bluetooth modules today?

Yes. The IOT747 IDC777 and IDC767 are fully integrated Dual Mode LE Audio and Classic Audio Bluetooth modules available through BCD Atlantik, suited to commercial audio products, hearing assistance devices, and industrial audio applications. Feasycom's BT 5.1 dual-mode audio modules are also available for applications requiring Classic Bluetooth audio alongside BLE. Contact us to discuss the right module for your LE Audio application.

Key takeaways

• Bluetooth 6.3 was adopted by the Bluetooth SIG in May 2026; it is a targeted set of enhancements to the 6.0 foundation, not a major generational shift in core wireless capability
• The four enhancements are: Channel Sounding Inline PCT Transfer (more accurate, lower-overhead ranging), PHY-specific RTT Accuracy (optimised ranging per PHY), HCI expansion (future-proofs the interface for new features), and ACP/C/I harmonisation (simplifies dual-mode radio design)
• For B2B applications requiring high-precision indoor positioning, asset tracking, or access control, 6.3's Channel Sounding improvements are the most operationally significant change
• For dual-mode products (Classic Bluetooth + BLE), the RF limit harmonisation reduces design complexity and opens the door to more efficient radio architectures
• Bluetooth 5.3 remains a fully capable, mature specification for the majority of B2B wireless connectivity applications; specifying 6.3 is only warranted where its specific capabilities address a real product requirement
• Certified 6.3 hardware is emerging through 2026; for programmes targeting 2026 production, 5.3 or 6.0 modules are the pragmatic specification; for 2027+ programmes, 6.3 is worth designing in from the outset

Conclusion

Bluetooth 6.3 represents a carefully targeted evolution of the specification rather than a reason to upgrade every B2B wireless product design. For engineers working in precision positioning, RTLS, or access control, the Channel Sounding improvements are material and worth planning for. For those designing dual-mode products or long-lifecycle platforms, the RF harmonisation and HCI expansion provide tangible design and future-proofing benefits. For the majority of B2B Bluetooth connectivity applications, Bluetooth 5.3 and 6.0 remain fully sufficient, and the practical question is when 6.3 certified modules will be available at the right specification for your specific product.

BCD Atlantik supplies Bluetooth modules from Feasycom, IOT747, Quectel, SparkLAN, and Flaircomm, covering BT 5.0 through to LE Audio capable and dual-mode audio modules. Browse our wireless connectivity range, or speak to our team to discuss the right Bluetooth module specification for your development programme and timeline.

See also:

How to choose the right IoT gateway | BCD Atlantik

What is Edge AI and why businesses are adopting it | BCD Atlantik

Raspberry Pi industrial applications have expanded steadily since the platform first appeared in engineering toolkits as a low-cost Linux development board. With the arrival of Raspberry Pi 5, that expansion has accelerated. The performance uplift, the first PCIe interface on a full-size Raspberry Pi, the rearchitected I/O silicon, and the growing ecosystem of HAT+ add-on boards - including dedicated AI accelerators - have shifted the Pi 5 from a capable prototyping platform into hardware that engineers are taking seriously for deployed industrial systems. This guide sets out what the Pi 5 offers, what its limitations are in industrial contexts, how it can be extended with the right add-on hardware, and where the Compute Module 5 is the more appropriate choice for production deployment.

Summary

The Raspberry Pi 5 is a 64-bit quad-core Arm Cortex-A76 single-board computer with significantly more processing power than its predecessor, a PCIe 2.0 interface for high-bandwidth peripherals, improved camera and display support, and a real-time clock. For industrial engineers, it is most valuable as a rapid prototyping and development platform, as a cost-effective edge compute node in benign environments, and as the foundation for AI-at-the-edge applications when paired with the AI HAT+ or AI HAT+ 2. For production deployment in harsh environments, the Compute Module 5 on an industrial carrier board remains the more appropriate choice.

What is the Raspberry Pi 5?

Raspberry Pi 5 is the fifth generation of the full-size Raspberry Pi single-board computer, and the first to use silicon designed in-house at Raspberry Pi. It is built around a 64-bit quad-core Arm Cortex-A76 processor running at 2.4GHz - the same CPU architecture as the Compute Module 5 - and is available in four RAM configurations: 2GB, 4GB, 8GB, and 16GB LPDDR4X.

The headline performance figure is a 2–3× increase in CPU performance relative to Raspberry Pi 4, which is significant in absolute terms and even more so in the context of what that headroom enables for edge computing and local data processing. But the specification changes that matter most to industrial engineers go beyond raw CPU speed.

Key specifications for industrial engineers

RP1 Southbridge and I/O Performance

For the first time, a full-size Raspberry Pi uses in-house silicon beyond the main SoC. The RP1 "southbridge" handles the bulk of I/O capability on the Pi 5, and the step change in peripheral performance it delivers is material for industrial use.

• USB bandwidth is more than doubled compared to Pi 4, enabling faster data transfer to external storage, industrial USB peripherals, and UAS drives
• The dedicated two-lane 1Gbps MIPI camera and display interfaces of previous models have been replaced by a pair of four-lane 1.5Gbps MIPI transceivers, tripling total bandwidth and supporting any combination of up to two cameras or displays
• Peak SD card performance is doubled through support for the SDR104 high-speed mode

For engineers building systems that involve camera feeds, industrial displays, or high-speed data ingest from USB peripherals, these are meaningful capability improvements rather than incremental refinements.

PCIe 2.0 x1 interface

The single-lane PCIe 2.0 interface exposed on the Pi 5 is the specification change with the most significant long-term implications for industrial and engineering use. PCIe enables the M.2 HAT+ adapter, which in turn enables NVMe SSD storage and AI accelerator modules to connect to the Pi 5 directly - providing the reliable solid-state storage and on-device inference performance that industrial edge applications increasingly require.

Real-Time Clock with external battery support

The on-board RTC, backed by an external battery connector, allows the Pi 5 to maintain accurate system time through power cycles and periods without network connectivity. For unattended industrial deployments where NTP synchronisation is not always available - remote monitoring nodes, equipment in electrically noisy environments, or deployments on intermittent cellular connectivity - this removes a practical gap that previous Pi models required third-party RTC add-on boards to address.

RAM and memory options

The 2GB variant suits lightweight IoT gateway and data logging applications; 4GB covers the majority of Linux-based industrial applications comfortably; 8GB is the practical choice for machine vision, concurrent process management, or applications running alongside a full desktop environment; 16GB is available for demanding edge AI inference workloads, particularly when running larger models locally. BCD Atlantik stocks the full range, with the 8GB variant the most commonly specified for industrial development work.

Extending Pi 5 for industrial and engineering applications

The HAT+ ecosystem is where the Pi 5's industrial utility expands most significantly. The PCIe interface that underpins the M.2 HAT+ opens up a hardware expansion path that was not available on previous full-size Pi models.

AI acceleration

Three AI accelerator products from Raspberry Pi are available through BCD Atlantik, all compatible with Pi 5 only via the PCIe interface.

The Raspberry Pi AI HAT+ offers 13 or 26 TOPS via a Hailo neural network accelerator, fitted directly to the Pi 5 via its PCIe interface. The 26 TOPS variant supports more complex neural processing and the ability to run multiple concurrent AI models - relevant for applications combining object detection with classification or tracking simultaneously.

The Raspberry Pi AI HAT+ 2 represents a substantial step up: 40 TOPS of INT4 inferencing performance via the Hailo-10H accelerator, with 8GB of dedicated on-board RAM. This dedicated RAM means the AI workload does not compete with the host Pi 5 for system memory, making the AI HAT+ 2 suitable for running large language models and vision-language models locally - enabling offline process control, secure local data analysis, facilities management automation, and advanced robotics applications that would otherwise require cloud connectivity for inference.

Storage expansion

The Raspberry Pi M.2 HAT+ connects M.2 2230 or 2242 format PCIe and NVMe devices to the Pi 5's PCIe connector, providing a path to fast, reliable NVMe SSD storage. The HAT+ specification allows Raspberry Pi OS to automatically detect the HAT+ and connected devices. For industrial applications where booting from an SD card is a reliability concern - and where the write endurance and data integrity guarantees of an industrial NVMe SSD are required - the M.2 HAT+ with an appropriate NVMe drive is the practical solution. BCD Atlantik also supplies SSD kits for Pi 5 with the SSD pre-assembled on the M.2 HAT+ for immediate use.

Power over Ethernet

The Raspberry Pi PoE+ HAT enables the Pi 5 to be powered directly from a PoE+ capable Ethernet switch, eliminating the need for a separate power supply at each deployment point. For distributed sensor networks, building automation nodes, or point-of-sale deployments where running separate power cabling to each device is impractical, PoE+ simplifies both installation and cable management significantly.

Environmental sensing

The Raspberry Pi Sense HAT 2 adds an array of onboard sensors - pressure, humidity, temperature, colour, orientation, and movement - alongside an 8×8 RGB LED matrix and five-button joystick. Originally developed for the International Space Station as part of the Astro Pi programme, it provides a compact, well-supported sensing platform for environmental monitoring, condition monitoring, and data logging applications where a single board can handle both sensing and compute.

Industrial applications of Raspberry Pi 5

• Rapid prototyping and system development: the Pi 5's combination of performance, standard Linux software stack, and accessible GPIO means engineering teams can prototype, validate, and iterate an industrial system design faster than on most proprietary embedded platforms
• Edge AI and machine vision: paired with the AI HAT+ or AI HAT+ 2, the Pi 5 becomes a capable on-device inference platform for quality inspection, anomaly detection, object counting, and camera-based monitoring applications
• Industrial IoT gateways: the Pi 5's networking capability, USB port count, and processing headroom make it well-suited to gateway applications aggregating data from multiple sensors or serial devices and forwarding it to a SCADA system or cloud platform
• HMI development and display systems: the improved MIPI display interface and GPU performance support the development of industrial HMI interfaces, digital signage, and kiosk systems on the standard Raspberry Pi OS desktop stack
• Laboratory and test equipment: data acquisition, instrument control, and automated test systems where the open ecosystem of Python libraries, GPIO access, and USB instrument connectivity makes the Pi 5 an accessible and capable platform
• Smart agriculture and environmental monitoring: distributed sensor nodes in agricultural, environmental, and infrastructure monitoring deployments where cost per node matters and benign operating conditions make standard Pi hardware appropriate

Limitations and considerations for industrial deployment

Engineers evaluating the Pi 5 for industrial use should be clear-eyed about where the standard hardware has constraints relative to purpose-built industrial computing platforms.

Operating temperature

The Raspberry Pi 5 is rated for operation between 0°C and 50°C ambient. Industrial environments regularly exceed this range - particularly in unventilated cabinets in summer, in vehicles, in outdoor enclosures, or in cold storage facilities. Applications in these environments require either active thermal management and an appropriately rated enclosure, or a purpose-built industrial computer such as the Edatec ED-IPC series based on the Compute Module 5, which is rated for -25°C to 60°C and available with optional 4G, isolated serial interfaces, and industrial power input.

Power input

The Pi 5 uses USB-C 5V/5A power delivery. Industrial deployments typically run on 9–36V DC from a DIN-rail power supply; there is no direct wide-voltage input on the standard Pi 5. Step-down converters can address this, but add board space and component count that an industrial carrier board design around the CM5 eliminates by design.

Storage reliability

Booting from an SD card remains the default for the Pi 5, and SD card reliability under continuous write workloads is a known concern for always-on deployments. The M.2 HAT+ with an industrial NVMe SSD resolves this, but adds cost and a HAT that must be accommodated in the physical design. For production deployments, this should be treated as a default rather than an optional upgrade.

Enclosure and mounting

The Pi 5 is a bare single-board computer; industrial deployment requires an appropriate enclosure, DIN-rail or panel mounting, and consideration of dust and moisture ingress. This is engineering effort that a purpose-built industrial carrier board or packaged computer eliminates.

Raspberry Pi 5 vs Compute Module 5: Choosing the right form factor

The same Raspberry Pi OS software stack runs on both platforms, which is the key practical advantage of this ecosystem: code developed and validated on a Pi 5 transfers directly to a CM5-based industrial carrier without a software bring-up process. This means the Pi 5 is the natural development and validation platform for applications that will ultimately be deployed on CM5-based hardware.

Best practices for Pi 5 in industrial and engineering projects

1. Replace SD card boot with NVMe from the start. For any deployment writing data continuously - logging, monitoring, vision systems - use the M.2 HAT+ with an industrial NVMe SSD as the default storage. SD card failure is the most common cause of unexpected Pi downtime in deployed systems.
2. Size RAM to the workload, not to the minimum. The cost difference between 4GB and 8GB is modest; the cost of an application running out of headroom in a deployed system is not. For anything involving concurrent processes or AI inference, specify 8GB as a baseline.
3. Match the AI accelerator to the inference requirement before specifying. The AI HAT+ (13 or 26 TOPS) suits single-model and concurrent vision tasks; the AI HAT+ 2 (40 TOPS, 8GB RAM) suits generative AI and vision-language model inference. Each has a distinct use case; choosing the wrong tier either under-specifies the application or adds unnecessary cost.
4. Validate thermal performance in the target enclosure, not on an open bench. The Pi 5 with an AI HAT+ under load generates meaningful heat; confirm that the Active Cooler is sufficient in the intended enclosure before committing to a physical design, particularly if convection cooling is limited.
5. Plan the transition to CM5 for production volume. If a project is likely to move from a handful of Pi 5 development units to production volume in an enclosure, plan the carrier board design early. The software is the same; the time cost is in the hardware design, not the OS bring-up.
6. Use the RTC battery connector from day one. Populating the RTC battery connector costs almost nothing and eliminates a class of time-synchronisation problems that affect unattended deployments, particularly those on intermittent network connectivity.

Frequently asked questions

Is the Raspberry Pi 5 suitable for 24/7 industrial deployment?

For deployments in controlled environments within its 0–50°C operating range, with NVMe storage rather than SD card, the Pi 5 can run continuously and reliably. For demanding environmental conditions - wider temperature ranges, DIN-rail cabinet installation, vehicle use - the Compute Module 5 on an industrial carrier board such as the Edatec ED-IPC series is the more appropriate platform.

What does the PCIe interface on Pi 5 enable that previous models couldn't do?

PCIe enables direct connection of NVMe SSDs (via the M.2 HAT+) and AI accelerator modules (AI HAT+ and AI HAT+ 2). Neither category of peripheral was directly connectable to previous full-size Pi models without workarounds. Both are significant for industrial and engineering applications.

Can the Raspberry Pi AI HAT+ 2 run LLMs locally?

Yes. The AI HAT+ 2's Hailo-10H accelerator and 8GB of dedicated on-board RAM are designed for running selected large language models and vision-language models locally on Raspberry Pi 5, without cloud connectivity. This enables offline process control, secure local data analysis, and conversational AI interfaces in air-gapped or connectivity-limited environments.

Does Raspberry Pi 5 support Power over Ethernet natively?

Not natively - PoE+ support requires the Raspberry Pi PoE+ HAT, which sits on the 40-pin GPIO header and draws power from the Ethernet port on a PoE+ capable switch. This is a well-supported configuration and removes the need for a separate USB-C power supply at each Pi 5 deployment point.

How does the Pi 5's MIPI camera support compare to previous generations?

Pi 5 replaces the dual two-lane 1Gbps MIPI interfaces of previous models with a pair of four-lane 1.5Gbps MIPI transceivers - tripling total bandwidth - and supporting any combination of up to two cameras or displays. Combined with the rearchitected Raspberry Pi Image Signal Processor, this makes Pi 5 significantly more capable for machine vision, multi-camera, and high-resolution display applications than its predecessors.

Can I use the same software on Pi 5 and CM5?

Yes. Both run the same Raspberry Pi OS on the same Arm Cortex-A76 architecture. Code, libraries, and application software developed on a Pi 5 will run on a CM5-based carrier board without modification, which makes the Pi 5 the natural development platform for systems that will be deployed on industrial CM5 hardware at production volume.

Is the Raspberry Pi 5 appropriate for safety-critical industrial applications?

The Raspberry Pi 5 and its software stack are not certified for functional safety applications (IEC 61508, ISO 26262, IEC 62061 etc.). For applications where failure could result in harm to people or equipment and where functional safety certification is a requirement, a certified industrial controller platform should be used. The Pi 5 is well-suited to monitoring, analytics, visualisation, and non-safety-critical control applications.

Key takeaways

• Raspberry Pi 5 delivers a 2–3× CPU performance uplift over Pi 4 on the same Arm Cortex-A76 architecture as the Compute Module 5, making the two platforms fully software-compatible
• The PCIe 2.0 interface is the most significant new capability for industrial engineers, enabling NVMe storage and AI accelerator add-on boards that were not possible on previous full-size Pi models
• The AI HAT+ (13 or 26 TOPS) and AI HAT+ 2 (40 TOPS, 8GB RAM, LLM/VLM capable) provide a tiered path to on-device AI inference suited to different application complexity levels
• The Pi 5 is best suited to prototyping, development, and controlled-environment deployment; the Compute Module 5 on an industrial carrier remains the right choice for production deployment in harsh environments
• Replacing SD card boot with NVMe SSD via the M.2 HAT+ is the single most important reliability improvement for any Pi 5 deployed in an always-on industrial context
• Software developed on Pi 5 transfers directly to CM5-based industrial hardware - making Pi 5 the natural development platform for the industrial embedded pipeline

Conclusion

Raspberry Pi 5 has arrived at a point where the platform's capabilities genuinely close the gap with dedicated industrial single-board computers for a broad range of engineering and industrial applications - not just in processing power, but in the I/O performance, storage options, and AI acceleration ecosystem that industrial deployments require. The addition of PCIe, the improved camera interfaces, and the HAT+ AI accelerator range have changed the nature of the conversation from "can the Pi do this?" to "at what scale and in what environment should it be doing this?"

For engineers building prototypes, developing edge AI applications, deploying low-volume systems in controlled environments, or validating software before a CM5-based production run, the Pi 5 is a more capable platform than any previous generation. Understanding where its limitations lie - and when the step to an industrial CM5 carrier board is warranted - is the practical knowledge this guide aims to provide.

BCD Atlantik is an authorised Raspberry Pi reseller supplying the full Pi 5 range in 2GB, 4GB, 8GB, and 16GB variants, alongside the AI HAT+, AI HAT+ 2, M.2 HAT+, PoE+ HAT, and SSD kits. We also supply the EDATEC range of CM5-based industrial computers for production-grade deployments. Browse our full Raspberry Pi range or contact our team to discuss your application.

See also:

What is Edge AI and why businesses are adopting it | BCD-Atlantik Ltd

Raspberry Pi Compute Modules for embedded systems | BCD-Atlantik Ltd

The edge computing vs cloud debate is one of the most practical architecture decisions B2B engineers face when designing connected systems in 2026. Both models have matured considerably over the past decade — cloud infrastructure has become more accessible and capable, while edge computing hardware has become compact, powerful, and affordable enough to deploy at scale outside of data centres. The question is no longer which one is better in principle, but which is the right fit for a given application, and how the two can work together as part of a coherent system architecture.

This guide cuts through the terminology to explain what each model is, how they differ technically and operationally, the benefits of edge computing for businesses alongside those of cloud computing, and how to approach the decision when designing or specifying an industrial or commercial connected system.

Summary

Cloud computing processes data in remote data centres and delivers results over the internet. Edge computing processes data locally, on hardware deployed at or near the source of the data. For B2B engineers, the core trade-off is latency and autonomy versus scale and flexibility: cloud computing offers virtually unlimited compute capacity and centralised management; edge computing offers millisecond-level response times, offline resilience, and reduced bandwidth costs. Most production deployments use both in combination, with edge handling real-time local decisions and cloud handling aggregation, storage, analytics, and model management.

What is cloud computing?

Cloud computing is the delivery of computing services — processing, storage, networking, databases, software — over the internet from shared data centre infrastructure operated by a third party. Rather than owning and maintaining physical servers, organisations rent capacity from a cloud provider and access it on demand.

For B2B applications, cloud computing typically provides:

• Scalable compute and storage that can be expanded or contracted to match workload demand without capital expenditure on physical hardware
• Centralised data aggregation, enabling an organisation to bring together data from many sources, sites, or devices into a single platform for analysis and reporting
• Managed software services — databases, machine learning platforms, messaging, dashboards — that reduce the development effort required to build a data pipeline
• Global accessibility, allowing authorised users to access systems and data from any location with an internet connection

The defining characteristic of cloud computing is that the data must travel to the processing resource — from the device or sensor, across a network, to a remote data centre — before a decision can be made or a result returned.

What is edge computing?

Edge computing brings processing capacity to the data source, rather than sending data to a centralised location for processing. An edge computing device — which might be an industrial computer, an embedded gateway, an edge AI station, or a ruggedised SoC-based module — sits physically close to the sensors, machines, or systems generating data and performs computation locally.

The "edge" in edge computing refers to the network edge: the boundary between local devices and the wider network. Rather than every data packet making a round trip to a cloud server, the edge device handles local processing, acts on the result immediately, and transmits only relevant or aggregated data onwards. The volume of data leaving the site, and the dependency on an external network, are both reduced significantly.

In a B2B context, edge computing hardware ranges from compact single-board industrial computers — such as the Edatec ED-IPC series, built on the Raspberry Pi Compute Module 5 and designed for DIN-rail deployment in industrial cabinets — to high-performance edge AI stations such as the Thundercomm TurboX EB5G2, which delivers 48 TOPS of on-device AI inference alongside 5G connectivity in a fanless industrial enclosure. At the silicon level, platforms such as the Qualcomm Dragonwing IQ-X Series package up to 45 TOPS of NPU performance into industrial-grade SoCs designed for integration into custom products, HMIs, and embedded controllers.

How they differ: Architecture and data flow

The fundamental architectural difference is where computation happens relative to the data source.

In a cloud-first architecture, a sensor or device collects data and sends it over a network — often via a cellular or wired connection — to a cloud platform. The cloud platform processes the data, stores it, and may return a result or trigger an action. Every decision requires a functioning network connection and introduces the latency of that round trip, which ranges from tens of milliseconds on a good connection to seconds or longer on a congested or intermittent one.

In an edge architecture, a local device receives the sensor data and processes it on-site. A decision is made and acted upon in milliseconds, without any dependency on network availability. Only the outputs — structured events, aggregated metrics, flagged anomalies — are transmitted onwards, either to a local SCADA system or to a cloud platform for longer-term storage and analysis.

In a hybrid architecture, both models operate in parallel. The edge handles real-time decisions; the cloud handles aggregation, historical analysis, model training, and reporting. This is the most common production pattern for mature IIoT and smart infrastructure deployments, because it uses each model where it is most suited rather than forcing one to do everything.

Edge computing vs cloud computing: Key differences at a glance

Benefits of edge computing for businesses

For B2B engineers evaluating where to put compute in a system design, the benefits of edge computing are most compelling in the following areas.

Low latency and real-time response

Any application where the output of an AI or logic decision must trigger an immediate physical action cannot tolerate a cloud round trip. A machine vision quality inspection system that needs to reject a defective part as it passes a camera at high speed, a safety monitoring system that must halt a machine when a person enters a hazard zone, or a process controller responding to a sensor threshold — all of these require decisions in milliseconds. Edge computing makes this possible; cloud computing does not.

Offline and degraded-network resilience

Industrial sites, remote infrastructure, vehicles, and temporary deployments frequently experience intermittent or unavailable network connectivity. An edge architecture continues to function and make decisions regardless of network state, buffering data locally and synchronising when connectivity is restored. A cloud-dependent system fails silently or noisily the moment the connection drops.

Reduced bandwidth cost and congestion

A multi-camera production line, a fleet of sensor-equipped vehicles, or a distributed infrastructure network can generate enormous volumes of raw data. Transmitting all of it to the cloud is expensive in terms of both bandwidth cost and network infrastructure. Edge computing filters and aggregates this data locally, transmitting only what is operationally relevant — reducing bandwidth requirements by orders of magnitude in many deployments.

Data privacy and sovereignty

In sectors including healthcare, defence, financial services, and critical national infrastructure, raw operational data may be subject to regulatory constraints on where it can be processed or stored. Edge computing keeps raw data on-site, reducing or eliminating the compliance exposure that arises from transmitting sensitive data to third-party cloud infrastructure.

Predictable and controllable infrastructure cost

Cloud compute costs scale with usage, offering flexibility for variable workloads but posing a risk for high-volume, continuous workloads where expenses can quickly escalate. In contrast, an edge device with a fixed hardware cost processes the same data volume for a consistent price, making cost modelling more predictable for operational deployments with a defined data volume.

Benefits of cloud computing for businesses

Cloud computing is not being displaced by edge — it remains the right model for a broad range of B2B workloads.

• Unlimited scalability: Cloud compute capacity can be expanded on demand, without physical hardware procurement or deployment, making it suited to applications with variable or unpredictable load profiles
• Centralised analytics and reporting: Aggregating data from many edge sites into a single cloud platform enables historical trend analysis, cross-site benchmarking, and management reporting that would be impractical to run on distributed edge hardware
• AI model training: Training machine learning models requires access to large datasets and significant GPU compute capacity, both of which are far more practical in the cloud than at the edge
• Global accessibility: Cloud-hosted applications and dashboards can be accessed by authorised users from any location, without requiring VPN access to a physical site
• Managed services and rapid development: Cloud platforms offer a broad ecosystem of managed databases, messaging systems, analytics tools, and APIs that reduce the development effort required to build a connected system

When to use edge, cloud, or both

The decision is rarely binary. A more useful framing is to identify which parts of a system's data flow are time-sensitive, connectivity-dependent, bandwidth-intensive, or privacy-sensitive, and apply each model to the parts of the architecture where it is best suited.

Use edge computing for: real-time control decisions, quality inspection, safety monitoring, local data pre-processing, offline-resilient systems, high-bandwidth sensor data filtering, and data that must not leave the site.

Use cloud computing for: historical storage and trend analysis, cross-site reporting, AI model training and management, global system access, and workloads with highly variable or unpredictable compute demand.

Use both in a hybrid architecture for: IIoT deployments where edge handles real-time site-level decisions and cloud handles aggregation, analytics, model distribution, and management; smart infrastructure where local autonomy is required but central oversight is also needed; and any application where the data volume is too large for full cloud transmission but the analytical value of aggregated data justifies centralised processing.

What edge computing hardware looks like in practice

For engineers new to deploying edge compute, the hardware landscape spans a wide range of form factors and performance levels. At the lower end, an industrial single-board computer such as the Edatec ED-IPC series - based on the Raspberry Pi CM5, DIN-rail mountable, with RS485/RS232 and optional 4G - provides a capable Linux edge node for protocol conversion, local data processing, and IoT gateway functions at modest cost and power draw.

For applications requiring higher AI inference throughput, edge AI stations such as the Thundercomm TurboX EB5G2 (48 TOPS, 5G, 24-channel HD video) or the TurboX EB6S (extensible up to 200 TOPS with AI accelerator cards) provide industrial-grade platforms suited to vision inspection, smart building analytics, and intelligent transport applications.

Where the edge compute capability needs to be built into a custom product, SoC-level platforms such as the Qualcomm Dragonwing IQ-X Series offer up to 45 TOPS of NPU performance in an industrial-grade SoC designed for integration into Windows-native industrial PCs and HMIs. The Dragonwing IQ6, IQ8, and IQ9 series address Linux-based embedded and IoT deployments across a range of performance tiers.

All of these represent different points on the same spectrum: the ability to process data where it is generated, act on it immediately, and reduce dependence on a network path that may be expensive, congested, or simply unavailable.

Best practices for a hybrid edge-cloud strategy

1. Define latency requirements first. If any part of the system requires a sub-100ms response to a sensor event, that part must run at the edge. This constraint should drive the architecture before any other consideration.
2. Audit connectivity assumptions. Map every site or deployment location against the network connectivity realistically available — not the best-case scenario. Sites that appear well-connected may have intermittent outages that a purely cloud-dependent architecture cannot tolerate.
3. Size edge compute to the workload, not to the maximum theoretical load. Over-specifying edge hardware adds cost without benefit; profile the actual inference or data processing workload and select hardware with appropriate headroom rather than maximum available TOPS.
4. Build remote management in from the start. A distributed fleet of edge devices without centralised remote management becomes an operational liability at scale. Platforms such as Thundercomm's OSware.Edge or Edatec's management tooling, alongside cloud-side orchestration, should be part of the architecture from day one.
5. Treat model and firmware updates as part of the system design. Edge AI deployments require a pipeline for updating models as they are retrained on new data. OTA (over-the-air) update capability should be specified and tested before deployment, not retrofitted after.
6. Consider data sovereignty requirements at the design stage. If the application involves data that must not leave a specific jurisdiction or physical site, the edge-cloud boundary in the architecture must enforce that constraint rather than relying on cloud provider settings that may change.

Frequently asked questions

Is edge computing replacing cloud computing?

No. Edge computing complements cloud computing rather than replacing it. The two models address different parts of a system's requirements. Cloud remains the most practical solution for large-scale storage, analytics, model training, and global access; edge provides the real-time, offline-resilient, bandwidth-efficient processing that cloud cannot deliver at the data source.

What types of B2B application benefit most from edge computing?

Applications with strict latency requirements (machine vision, safety systems, real-time control), deployments in locations with limited or intermittent connectivity (remote infrastructure, vehicles, temporary sites), and use cases involving sensitive data that must not leave the site (healthcare, defence, financial services) tend to benefit most from edge compute. High-bandwidth sensor environments where transmitting all raw data to the cloud is cost-prohibitive also suit an edge-first approach.

What does edge computing hardware typically cost for a B2B deployment?

This varies considerably by performance tier. A compact industrial edge computer based on the Raspberry Pi Compute Module can cost from a few hundred pounds per unit. A high-performance edge AI station with 5G connectivity and multi-camera support typically costs between £500 and £3,000 depending on the AI accelerator configuration. SoC and SOM-based solutions for custom product integration are priced differently again, typically sold in volume for integration into a manufactured product.

Can edge and cloud architectures share the same data platform?

Yes, and this is the standard pattern for mature deployments. Edge devices process data locally and transmit structured events or aggregated telemetry to a cloud platform, which handles storage, visualisation, alerting, and cross-site analytics. Platforms such as AWS IoT Greengrass and Azure IoT Edge provide explicit support for this hybrid model, with local runtime capability on the edge device and cloud-side management and analytics.

How is edge computing managed at scale?

Managing a fleet of edge devices requires dedicated tooling for remote configuration, firmware updates, health monitoring, and model deployment. Most industrial edge computing platforms include or integrate with a device management layer — Thundercomm's OSware.Edge and edgeOS platforms, for example, provide OTA update, remote monitoring, and edge-cloud synchronisation features specifically designed for fleet management of deployed edge devices.

What connectivity does an edge computing device need?

This depends on the application. A fully local edge deployment may need only a connection to the local sensor network or machine interface, with periodic connectivity for data sync. A deployment requiring real-time cloud integration needs reliable WAN connectivity, typically via Ethernet, Wi-Fi, or cellular (4G/5G). Industrial 5G gateways, such as the Teltonika RUTX50, are commonly deployed alongside edge computing devices at sites where wired connectivity is unavailable.

Is edge computing suitable for small and mid-sized businesses?

Yes, particularly at the lower end of the hardware spectrum. Compact industrial edge computers are cost-effective enough that SMEs in manufacturing, agriculture, retail, and building management are deploying them at individual sites or in small fleets. The availability of open-source frameworks and managed edge platforms has also reduced the software development effort required to deploy a functional edge system.

Key takeaways

• Cloud computing processes data in remote data centres; edge computing processes data locally at the source — the choice between them is primarily about latency, connectivity, bandwidth, and data sovereignty requirements
• Neither model is universally superior; most production B2B deployments use both in a hybrid architecture where edge handles real-time local decisions and cloud handles aggregation, analytics, and management
• The core benefits of edge computing for businesses are low latency, offline resilience, reduced bandwidth cost, and improved data privacy — each most relevant in specific application contexts
• Edge computing hardware spans a wide range, from low-cost industrial single-board computers to high-performance AI stations and industrial-grade SoCs with tens of TOPS of on-device inference capability
• Remote management, OTA updates, and a clear edge-cloud data pipeline should be designed into any edge deployment from the start, not treated as later additions
• Latency requirements, connectivity constraints, data sovereignty obligations, and bandwidth costs are the four most useful inputs to an edge-vs-cloud architecture decision

Conclusion

For B2B engineers, the edge vs cloud question is most productively framed not as a competition but as a resource allocation problem: given the latency, connectivity, bandwidth, and privacy constraints of a specific application, where should each part of the data processing happen? Cloud computing offers scale and flexibility that edge hardware cannot match; edge computing offers speed, resilience, and data control that a cloud-dependent architecture cannot provide.

Understanding both models and how they interact is increasingly a foundational skill for engineers designing connected industrial and commercial systems, as the volume of data generated at the edge continues to outpace what it is practical or cost-effective to transmit to the cloud.

BCD Atlantik supplies edge computing hardware from Thundercomm, Edatec, and Qualcomm Dragonwing, alongside connectivity hardware from Teltonika and Lantronix to support the network infrastructure edge deployments depend on. Browse our edge computing and connectivity range, or speak to our team about the right hardware for your application.

Flexxon X-PHY AI SSD security represents a different philosophy from conventional endpoint protection. Rather than adding another layer of software above the operating system, the X-PHY embeds an AI security engine directly into the SSD firmware — operating below the OS, below any application, and below the reach of most attack vectors that software-based defences cannot intercept. First launched in 2021 as the world's first AI-embedded cybersecurity solid-state drive, the X-PHY has since attracted attention from sectors where data breaches carry the highest consequences: healthcare, industrial operations, government, and financial services. This guide explains how the technology works, what it protects against, why hardware-level security addresses threat categories that software cannot, and what organisations evaluating endpoint security for high-sensitivity environments need to understand before specifying it.

Summary

The Flexxon X-PHY is a PCIe NVMe SSD with an AI security engine embedded in its firmware that monitors data access patterns at the NVMe protocol level, around the clock, without relying on the host operating system. When it detects access patterns consistent with ransomware, malware, physical tampering, or drive theft, it autonomously locks the drive and alerts the user - without waiting for human intervention or a software update. It is designed to function as a last line of defence: a security layer that continues to operate even if every software-based protection on the device has been compromised or bypassed.

What is the Flexxon X-PHY?

The X-PHY is a PCIe NVMe 1.3 M.2 2280 solid-state drive with capacities from 512GB to 1TB, built on 3D NAND flash with LDPC error correction, dynamic and static wear levelling, power loss protection, and SMART health monitoring - specification characteristics consistent with an industrial-grade drive. What distinguishes it from any other NVMe SSD is the presence of Flexxon's proprietary AI One Core Quantum Engine in the firmware alongside the standard storage controller.

This AI engine is not a separate module, an add-in card, or a software agent. It is embedded directly into the SSD firmware, operating at the same level as the storage controller itself. It has no dependency on the host OS, no requirement for signature updates, and no external connection. From the moment the drive powers on, it is monitoring.

The X-PHY is compatible with Windows and Linux, connects via M.2 or USB adapter, and is configured through Flexxon's Windows or Linux application. Once configured, it operates autonomously. More detailed information and the complete Flexxon product range are available from BCD Atlantik.

How the X-PHY AI engine works

The AI One Core Quantum Engine operates at the NVMe protocol layer - monitoring the data access patterns between the host system and the SSD in real time. Every read and write request that passes through the NVMe interface is analysed against a behavioural model of what normal, legitimate access looks like for that device and workload.

This is a fundamentally different detection methodology from software antivirus or endpoint detection and response (EDR) tools, which operate at the application or OS level and look for known threat signatures or behavioural indicators in running processes. The X-PHY works at a lower level: it is watching the actual data traffic to and from the storage medium, in the specific NVMe protocol that storage devices communicate in. Ransomware, for example, exhibits characteristic read/write patterns as it enumerates and encrypts files; the X-PHY identifies those patterns as anomalous and acts before the encryption completes, without needing to know the specific ransomware variant or having seen it before.

When a threat is detected, the X-PHY autonomously locks the drive and simultaneously alerts the user. Access can only be restored through authentication via the X-PHY Connect Bluetooth application or the Windows management tool, using two-factor authentication. In extreme cases, a rapid purge capability allows the drive's contents to be remotely wiped to prevent unauthorised access to the stored data.

The engine is self-learning: it builds and refines its model of normal access behaviour over time, improving its ability to distinguish between legitimate and anomalous access patterns without requiring manual reconfiguration or signature updates.

X-PHY security feature set

Flexxon describes the X-PHY's protection across five distinct security capabilities.

X-Guard Threat Lock

Real-time protection against malware, ransomware, and virus activity through continuous monitoring of read/write patterns. When anomalous patterns consistent with malicious data access are detected, the drive locks autonomously, immediately halting any ongoing attack on the stored data without waiting for a human decision or a software response.

X-Stream Protection

Security of the signal communication path between the host and the SSD, including inter-module data communication encryption. This addresses attack vectors that target the data in transit between the processor and the storage medium rather than the stored data itself.

X-File Forensic Agent

Audit trail generation for data access events, providing a verifiable record of what accessed the drive, when, and with what result. For organisations operating under regulatory frameworks that require demonstrable data access logging — healthcare under NHS data security standards, financial services under FCA requirements, or government systems under relevant security classifications — this creates an evidence layer that software-level logging cannot provide with the same integrity guarantees, since it operates below the OS where logs could otherwise be manipulated.

X-Site Physical Attack Protection

Hardware sensors embedded in the SSD detect physical attack vectors including drive removal, chip top-layer scraping, cloning attempts, side-channel attacks using power surges, and anomalous temperature fluctuations that may indicate physical tampering with the device. When any of these conditions is detected, the drive enters immediate lockdown. This protection class is largely absent from software-only security approaches, which have no mechanism to detect or respond to physical interference with the storage device.

X-Factor Encryption Lock

Protection of the security key itself and the SSD firmware against attack, preventing adversaries from targeting the encryption layer or the drive firmware as an alternative route to the stored data. Firmware attacks have become an increasingly favoured vector for sophisticated threat actors, precisely because most endpoint security software does not monitor or protect at the firmware level.

Why hardware-level security addresses gaps that software cannot

The case for hardware-embedded security at the storage layer rests on a structural limitation of software-based approaches. Antivirus, EDR, and even full-disk encryption all depend on the integrity of the operating system and the software stack running on top of it. If an attacker gains sufficient privilege on the host - through a zero-day vulnerability, a compromised driver, a supply-chain attack, or insider access - they can disable, bypass, or manipulate software security tools before those tools can respond.

The X-PHY is not accessible to the host OS in the same way a software agent is. Its AI engine operates in the SSD firmware, a layer the OS cannot directly modify or disable. The X-PHY operates and detects within the specific NVMe protocol, while software detects threats in an external open environment where there are many attack vectors — a narrower, more controlled monitoring context that is significantly harder for an attacker to spoof or evade.

This matters particularly for zero-day threats. Antivirus protection requires constant updating of the signature library and relies heavily on human intervention, while the X-PHY technology eliminates the need for human intervention and provides 24/7 threat protection against any new threats, including zero-day exploits. A threat that has never been seen before, and therefore has no antivirus signature, is still identifiable through the anomalous access patterns it produces at the storage layer.

Physical threat protection represents a further gap that software cannot address at all. The X-PHY features protection against physical incursions, such as drive detachment and cloning attacks, along with rapid purge capabilities — capabilities that require hardware sensors embedded in the device itself.

X-PHY vs conventional endpoint security approaches

Benefits for B2B organisations

• Protection that persists when the OS is compromised: The X-PHY continues operating and enforcing access controls even if the host system has been fully compromised by malware, ransomware, or a privileged attacker
• Zero-day ransomware coverage without signature updates: Behavioural detection at the NVMe level means unknown ransomware variants trigger lockdown based on their access patterns, not their identity
• Physical security for mobile and field-deployed devices: Laptops, rugged tablets, and portable equipment that may be lost, stolen, or physically tampered with benefit from drive-level sensors that no software tool can replicate
• Audit-ready forensic logging: The X-File Forensic Agent provides tamper-resistant access records that software-level logging cannot guarantee, supporting regulatory compliance and incident investigation
• Simplified compliance posture: A hardware-layer last line of defence reduces reliance on consistent user behaviour and software update discipline — both of which are significant sources of vulnerability in organisations operating at scale
• Minimal operational overhead: Once configured, the X-PHY operates autonomously without requiring signature updates, policy tuning, or ongoing management attention beyond standard device management

Applications and use cases

Healthcare and medical devices

Patient data is among the most heavily regulated and most frequently targeted categories of sensitive information. Medical devices increasingly store and process clinical data on embedded or attached storage; a breach or ransomware attack on a device in a clinical environment carries both regulatory consequences and direct patient risk. The X-PHY's combination of autonomous ransomware detection, physical attack protection, and forensic logging suits healthcare endpoint security requirements where regulatory accountability and operational continuity are both critical.

Industrial and operational technology

As OT networks converge with IT infrastructure, industrial endpoints — HMIs, SCADA workstations, industrial PCs — become accessible to the same threat actors targeting enterprise IT, but often run older operating systems with limited software security support. The X-PHY operates independently of the host OS, making it relevant for industrial endpoints where the OS cannot be updated or where software agents cannot be deployed without certification risk.

Government and defence

Government systems and defence endpoints handling sensitive or classified data face a combination of sophisticated nation-state threat actors and strict physical security requirements. The X-PHY's physical tamper detection and rapid purge capability directly address the threat of device seizure or physical attack on storage, complementing software-level controls that cannot operate after physical compromise.

Financial services

Financial services organisations are subject to regulatory requirements around data integrity, access logging, and incident response. The X-File Forensic Agent's hardware-level audit trail and the X-PHY's autonomous response to anomalous access patterns support both compliance requirements and the incident response capability that regulators increasingly expect.

Legal, professional services, and intellectual property

Organisations holding high-value confidential data — legal firms, engineering consultancies, product developers — face insider threat risk as well as external attack. The X-PHY uses the Zero Trust security framework to monitor for threats 24/7 in real-time, and 68% of organisations are vulnerable to insider attacks. Hardware-level monitoring of data access patterns provides a detection capability that an insider threat cannot disable through application-level manipulation.

Deployment and integration

The X-PHY connects via M.2 slot (PCIe NVMe 1.3, M.2 2280 form factor) or via USB adapter, making it compatible with laptops, desktop workstations, and a range of industrial computing platforms. Initial configuration is performed through Flexxon's Windows or Linux management application, which allows administrators to set security policies, review the forensic audit trail, and manage unlock credentials.

Day-to-day operation is autonomous: the drive monitors and responds without generating constant management overhead. When a lockdown event occurs, the assigned administrator or user is alerted and authenticates to restore access via the X-PHY Connect Bluetooth application or the Windows tool, using two-factor authentication.

The X-PHY Guard Solution, the latest iteration of the platform, is available in Lite, Essential, and Premium editions with an annual subscription, providing access to ongoing AI engine updates and the management platform.

The X-PHY is intended to complement existing security infrastructure — antivirus, firewall, EDR, network monitoring — rather than replace it. The value proposition is the security layer it adds below the point where conventional tools operate, closing the gap between what software can protect and what a determined or sophisticated attacker can bypass.

Frequently asked questions

Is the Flexxon X-PHY a replacement for antivirus or EDR software?

No. The X-PHY is designed as a last line of defence that operates at a layer below conventional security software. It complements antivirus, EDR, firewall, and other controls rather than replacing them. Its value is in protecting data at the storage level when upper-layer security has been bypassed or compromised.

How does the X-PHY detect ransomware it has never seen before?

The AI engine monitors data access patterns at the NVMe protocol level rather than looking for known signatures. Ransomware produces characteristic read/write behaviour as it enumerates and encrypts files; the X-PHY identifies this pattern as anomalous and locks the drive regardless of the specific ransomware variant involved. This behavioural approach means previously unknown threats are detectable without any signature update.

What happens to data if the drive locks down?

Data on a locked X-PHY drive is inaccessible until an authorised user authenticates to unlock it, via the X-PHY Connect Bluetooth application or the Windows management tool with two-factor authentication. The data is not destroyed by a lockdown event unless the administrator explicitly triggers the rapid purge function, which is available for scenarios where the device may be in the hands of an unauthorised party.

Can the X-PHY be disabled by malware running on the host?

The AI engine operates in the SSD firmware, below the operating system. Malware running at the OS or application level cannot directly access or disable the firmware-level security engine, which is a key advantage over software-based endpoint security tools that a sufficiently privileged attacker can disable or circumvent.

Which form factors and interfaces does the X-PHY support?

The X-PHY is available in PCIe NVMe M.2 2280 form factor, and is also accessible via USB adapter. It is compatible with Windows and Linux operating systems, covering the majority of enterprise and industrial endpoint platforms.

What is the X-PHY Guard Solution and how does it differ from the original X-PHY?

The X-PHY Guard Solution is the current platform iteration, available in Lite, Essential, and Premium subscription tiers. It combines the AI-embedded SSD hardware with an annual service subscription that provides ongoing AI engine updates, access to the management platform, and monitoring capabilities. Contact BCD Atlantik for current pricing and edition details.

Is the X-PHY suitable for use with industrial computers and edge devices?

Yes, wherever the host device includes an M.2 NVMe slot. Industrial computers such as those in the Edatec IPC series support M.2 NVMe SSD expansion; the X-PHY can be used as the primary or secondary storage device in such systems, adding hardware-level security to edge and embedded deployments where software security tooling may be limited or unavailable.

What sectors is the X-PHY most commonly deployed in?

Flexxon positions the X-PHY primarily for healthcare, industrial, government, and financial services deployments — sectors where the regulatory, operational, or safety consequences of a data breach are highest and where the limitations of software-only security are most material. It has also been integrated by Lenovo into a range of laptops for enterprise and high-security user deployments.

Key takeaways

• The Flexxon X-PHY embeds an AI security engine in SSD firmware, operating below the OS — a layer conventional antivirus, EDR, and encryption tools cannot reach and attackers cannot easily disable
• Threat detection is behavioural rather than signature-based, enabling autonomous protection against zero-day ransomware and unknown malware without requiring signature updates or human intervention
• Hardware sensors provide physical attack protection — detecting drive removal, cloning, chip scraping, power surges, and temperature anomalies — a capability entirely absent from software-based endpoint security
• The X-File Forensic Agent generates a hardware-level audit trail that cannot be manipulated at the OS level, supporting regulatory compliance and incident investigation requirements
• The X-PHY is designed as a last line of defence complementing existing security controls, not a replacement for them; its value is the gap it closes below the layer where software security operates
• Use cases span healthcare, industrial OT, government, financial services, and any environment where endpoint data integrity is critical and physical security of storage devices cannot be guaranteed

Conclusion

The Flexxon X-PHY addresses a specific and growing problem in endpoint security: the assumption, implicit in most conventional security architectures, that if an attacker reaches the OS they can also reach everything the OS can reach. For environments where that assumption carries unacceptable risk — where data must be protected even if the host system is fully compromised, physically stolen, or subjected to a sophisticated firmware attack — hardware-embedded AI security at the storage layer provides a protection class that software alone cannot deliver.

For organisations evaluating endpoint security for high-sensitivity deployments, the X-PHY is not a product that replaces a security programme. It is the component that continues operating when everything else has failed.

BCD Atlantik is an authorised Flexxon distributor, supplying the X-PHY AI Cyber Secure SSD alongside Flexxon's full range of industrial memory, WORM storage, and cybersecure storage solutions. Browse the full Flexxon range at BCD Atlantik, or contact our team to discuss the right specification for your endpoint security requirements.

Edge AI for business, explained simply: it is the ability to run artificial intelligence workloads directly on a device at the point where data is generated; on a factory floor, inside a vehicle, at a retail checkout, or on a piece of infrastructure equipment - rather than sending that data to a remote cloud server for processing. In 2026, edge AI has moved from being a competitive advantage for early adopters to a practical operational tool that B2B companies across manufacturing, logistics, healthcare, retail, and infrastructure are actively deploying.

This guide explains what edge AI is, why it behaves differently from cloud-based AI, what is driving adoption right now, and what the technology actually looks like when it is implemented in a business environment.

Summary

Edge AI is AI inference running locally on hardware at or near the data source, rather than in a centralised cloud. It matters for business because it eliminates the latency, bandwidth cost, and data privacy exposure that come with sending raw data offsite for processing. The result is faster decisions, lower connectivity costs, improved data security, and AI capability that works even when network connectivity is unavailable or intermittent. Adoption is accelerating because the hardware to do this has become compact, power-efficient, and cost-effective enough to deploy at scale outside of controlled data centre environments.

What is edge AI?

To understand edge AI, it helps to separate two concepts that are often used interchangeably: artificial intelligence and inference.

Training is the computationally intensive process of building an AI model - feeding it large datasets and adjusting millions of parameters until it can reliably identify patterns, classify inputs, or make predictions. Training typically happens in the cloud or a data centre, on large clusters of GPUs. It is expensive, slow, and power-hungry.

Inference is the process of using a trained model to make a decision or prediction on new data. Inference is what actually happens in deployment: the model sees a camera frame and identifies a defect; it reads a sensor signal and predicts bearing wear; it analyses a transaction and flags anomalous behaviour. Inference is far less computationally demanding than training, and it is inference that edge AI hardware is designed to run - quickly, efficiently, and locally.

Edge AI, therefore, is inference at the edge: a trained model, compressed and optimised for efficient execution, running on a device deployed in the field. The model might have been trained in the cloud; the decisions it makes happen locally, in real time, without a round trip to any external server.

Edge AI vs Cloud AI: What is the difference?

Understanding the distinction helps clarify why businesses are choosing to deploy AI at the edge rather than, or alongside, the cloud.

Neither model is inherently superior. Many B2B deployments use both: edge AI handles real-time, latency-sensitive decisions locally, while aggregated insights and model updates flow to and from the cloud on a lower-frequency basis. The edge and cloud complement each other rather than compete.

How edge AI works in practice

An edge AI deployment typically involves three components: a sensor or data source (a camera, a machine vibration sensor, a barcode scanner, a microphone), an edge AI device running inference, and some form of output - an alert, a control signal, a log entry, or a dashboard update.

A practical example: A food packaging line uses a camera to inspect each item before it is sealed. Historically, that inspection was done by a human or by a rules-based machine vision system. With edge AI, a trained neural network runs directly on an edge device mounted on the production line. It analyses each frame in real time, identifies defects, and triggers a reject mechanism in under a millisecond - far faster than a cloud round trip, and without sending a continuous HD video feed offsite. The model was trained on thousands of images of acceptable and defective product; the inference that happens on the line costs a fraction of a watt of additional compute.

The edge AI device in this scenario might be a compact embedded computer with a dedicated neural processing unit (NPU) - a specialised processor optimised for the matrix arithmetic that underpins neural network inference. NPU performance is measured in TOPS (Tera Operations Per Second), a metric that has become the standard shorthand for comparing edge AI compute capacity across hardware platforms.

What is driving B2B adoption in 2026?

Several converging factors have accelerated the business case for edge AI beyond where it stood even two or three years ago.

Hardware has become fit for industrial deployment

The NPUs that make edge AI inference practical have migrated from smartphone SoCs into purpose-built industrial platforms. Qualcomm's Dragonwing IQ series, for example, extends the company's AI processing architecture - the same Hexagon NPU found in Snapdragon-powered devices - into industrial-grade SoCs rated for extended temperature ranges and long product lifecycles. The IQ-X Series targets Windows-native industrial PCs and HMIs, with a Hexagon NPU rated up to 45 TOPS alongside Oryon CPU cores, designed for factory automation and edge AI vision applications. The IQ6, IQ8, and IQ9 series target Linux-based embedded and IoT deployments across a range of performance and power envelopes.

At the system level, edge AI stations such as the Thundercomm TurboX EB5G2 (48 TOPS, powered by the Qualcomm QCS8550 on a 4nm process) and the TurboX EB6S (15 TOPS base SoC, expandable to 70 or 200 TOPS via AI accelerator card) package this capability into fanless industrial computers with 5G connectivity, multiple camera inputs, and industrial interfaces - ready to deploy in a factory, a retail unit, or a vehicle, rather than in a rack in a climate-controlled server room.

Model compression has made inference tractable on compact hardware

Techniques such as quantisation (reducing the numerical precision of model weights from 32-bit to 8-bit or lower), pruning, and knowledge distillation have made it possible to run models that would previously have required a GPU server on hardware consuming a few watts. A model that achieves near-identical accuracy to a large cloud model but runs in 10 milliseconds on a compact edge device is a fundamentally different proposition for operational deployment.

Regulatory and data sovereignty pressure

Industries handling sensitive data - healthcare, financial services, critical national infrastructure - face growing regulatory expectations around where data is processed and stored. Edge AI is an architectural response to these constraints: if patient images, biometric data, or operational data never leave the local device, the regulatory exposure associated with transmitting them to a cloud provider is eliminated.

Connectivity economics

A multi-camera smart factory generating continuous HD video from fifty production lines produces data volumes that are prohibitively expensive to transmit to the cloud at full fidelity. Edge AI changes the economics: instead of transmitting raw video, the edge device transmits only structured events - "defect detected, Line 3, Camera 7, 14:23:07" - reducing bandwidth consumption by orders of magnitude while retaining all the operationally relevant information.

B2B applications of edge AI

Manufacturing and quality control

Machine vision for automated defect detection, tool wear monitoring, assembly verification, and worker safety compliance are among the most mature edge AI applications. The low latency of on-device inference is essential for any application where the AI output must trigger an immediate physical response - ejecting a defective product, stopping a machine, or raising an alert before a dangerous condition develops.

Smart retail

Customer behaviour analytics, shelf availability monitoring, queue management, and loss prevention using on-device video analytics. Edge AI allows retailers to extract operational intelligence from camera systems without the cost and compliance complexity of streaming footage offsite.

Healthcare and medical devices

Diagnostic support, patient monitoring, and medical imaging analysis benefit from edge AI where patient data cannot leave the clinical environment, or where real-time response is required independently of network connectivity. Edge inference on medical imaging devices is an area of growing development activity.

Transportation and logistics

Vehicle-mounted AI for driver monitoring, load inspection, route optimisation, and predictive maintenance. Platforms such as the Thundercomm EB5G2, with integrated 5G and multiple camera interfaces, are well suited to in-vehicle edge AI deployments where connectivity is intermittent and latency requirements are strict.

Smart infrastructure and utilities

Substation monitoring, pipeline anomaly detection, traffic management, and smart grid applications use edge AI to process sensor data locally and act on it without depending on a centralised control system. At remote sites with limited connectivity, the ability to operate autonomously is often a baseline requirement rather than a feature.

Security and surveillance

On-device object detection, people counting, and anomaly recognition in video feeds, replacing rules-based motion detection with AI-driven analysis that can distinguish between a genuine security event and an innocuous movement. Processing locally means lower storage requirements and reduced privacy exposure from continuous cloud video streaming.

What to look for when evaluating edge AI hardware

For businesses beginning to evaluate edge AI hardware, the following factors tend to define the practical fit between a platform and an application.

• TOPS rating and the model it needs to run: TOPS is a useful comparative metric but is not the only one that matters. The NPU architecture, supported precision levels (INT8, INT4, FP16), and compatibility with common AI frameworks (ONNX, TensorFlow Lite, PyTorch Mobile) determine whether a given model will run efficiently on a given platform
• Camera and sensor input support: multi-camera deployments require ISP capability and sufficient interface bandwidth; platforms like the Qualcomm QCS8250 support up to 24 simultaneous camera feeds, which matters considerably for larger surveillance or inspection deployments
• Industrial connectivity: 5G, Wi-Fi 6, Ethernet, and industrial fieldbus interfaces (RS232/RS485/CAN) determine how the edge device integrates into the existing site infrastructure
• Software and model deployment tooling: edge AI stations with built-in management platforms (such as OSware.Edge on the Thundercomm EB series) simplify OTA model updates, device monitoring, and multi-site deployment without custom infrastructure development
• Environmental rating: fanless industrial design, wide operating temperature, and DIN-rail or panel mounting options determine whether the platform can be deployed in the intended environment without additional enclosure engineering
• Scalability path: a platform that supports AI accelerator expansion (such as the EB6S with its optional 70 or 200 TOPS accelerator cards) allows deployments to grow compute capacity without replacing the base hardware

Frequently asked questions

Do I need internet connectivity to run edge AI?

No. That is one of edge AI's primary advantages. Once a model is deployed on the device, inference runs entirely locally. Connectivity may be used for model updates, telemetry transmission, or cloud integration, but is not required for the AI workload itself.

What is a TOPS rating and how much do I need?

TOPS (Tera Operations Per Second) measures the number of arithmetic operations a processor can perform per second at a given numerical precision. Higher TOPS allows more complex models or more simultaneous inference tasks. A single-camera quality inspection application might run comfortably on 10–15 TOPS; a multi-camera smart factory deployment running several concurrent models simultaneously may require 50 TOPS or more. Requirements should be derived from the specific models and workloads involved rather than from TOPS figures alone.

How is an edge AI model different from a cloud AI model?

The underlying neural network architecture may be identical, but edge-deployed models are typically compressed and quantised for efficient execution on constrained hardware. This involves some accuracy trade-off that must be validated for the specific application, though for many industrial use cases the practical difference is negligible.

Can edge AI hardware be updated as models improve?

Yes. Most industrial edge AI platforms support over-the-air (OTA) model updates, allowing improved models to be pushed to deployed devices without physical access. This is a key operational requirement for real-world deployments where retraining and model improvement are part of an ongoing operational process.

Is edge AI only relevant for large enterprises?

No. The cost profile of edge AI hardware has fallen significantly, and compact, lower-TOPS platforms are accessible to smaller manufacturers and operators. The operational benefits - reduced cloud costs, improved latency, offline resilience - are relevant at almost any scale.

What is the difference between an edge AI station and an edge AI SoC or SOM?

An edge AI station (such as the Thundercomm TurboX EB series) is a complete, deployable device with enclosure, power management, cooling, and interfaces included. A SoC (System on Chip) or SOM (System on Module) such as the TurboX C6490 or Qualcomm Dragonwing IQ series is a component or module that an engineering team integrates into a custom carrier board or product design. Edge AI stations suit deployment of AI capability in existing operations; SoCs and SOMs suit building AI capability into a new product.

How does edge AI affect data privacy and compliance?

Significantly and positively, in most cases. When inference runs locally and only structured results (rather than raw video, audio, or biometric data) leave the device, the data governance and compliance obligations associated with transmitting sensitive data to third-party cloud infrastructure are reduced or eliminated. This is a meaningful advantage in healthcare, financial services, and public sector deployments.

Key takeaways

• Edge AI is AI inference running locally on hardware at the data source, rather than in a remote cloud - faster decisions, lower bandwidth cost, better privacy, and offline resilience
• Training still happens in the cloud; edge AI is about deploying the resulting models where the data is generated
• Three factors are driving B2B adoption in 2026: industrial-grade hardware has become available at deployable cost and form factor; model compression techniques have made powerful inference tractable on compact devices; and regulatory and connectivity economics favour local processing for many applications
• Manufacturing quality control, smart retail, healthcare, transport, and infrastructure are the leading sectors for edge AI deployment today
• Evaluating edge AI hardware requires looking beyond TOPS to framework compatibility, connectivity, environmental rating, and model deployment tooling
• Edge AI stations suit operational deployment; SoCs and SOMs suit integration into custom product designs - both have their place in a B2B technology strategy

Conclusion

Edge AI is not a future technology. It is being deployed today by businesses that need decisions faster than a cloud round trip allows, data privacy that cloud processing cannot guarantee, and AI capability that does not fail when the internet connection does. The hardware that makes this possible - compact, fanless, industrially rated, with enough on-device compute to run meaningful AI models - has reached the point where it can be deployed outside a controlled lab environment and into real-world industrial and commercial operations.

For businesses at the beginning of their edge AI journey, the most valuable first step is usually identifying the specific decisions that need to be made faster or more reliably than current processes allow, and working backwards to the hardware and model requirements from there.

BCD Atlantik supplies edge AI hardware from Thundercomm and Qualcomm Dragonwing, ranging from deployable edge AI stations to industrial SoCs and SOMs for embedded product development. Browse our edge computing range, or speak to our team about the right starting point for your application.

Choosing the right 5G router for business in 2026 is no longer a niche decision reserved for large enterprises or telecoms specialists. UK 5G coverage has expanded significantly - Ofcom data from late 2025 places EE at 89% outdoor premises coverage, O2 at 76%, Three at 71%, and Vodafone at 64% - which means the connectivity is now available across a wide enough footprint to make 5G a credible primary or failover WAN option for most business sites. The challenge has shifted from availability to selection: which router is the right fit for your application, your environment, and your management requirements?

This guide covers what separates a genuine business-grade 5G router from a consumer device, the specifications that matter most at the evaluation stage, and a comparison of the options available from BCD Atlantik to help you narrow down your shortlist.

Summary

For most UK businesses in 2026, the right 5G router comes down to three questions:

• Is this a fixed-site deployment or a mobile/vehicle application?
• Do you need integrated enterprise security and cloud management, or a more open and configurable platform?
• And how important is 4G fallback coverage for locations where 5G signal is inconsistent?

The Teltonika RUTX50 suits industrial and IoT-adjacent fixed or semi-fixed deployments where configurability and remote management matter; the Lantronix G520 suits enterprises and critical infrastructure operators where built-in security, fieldbus protocol support, and ConsoleFlow cloud management are priorities.

What is a business 5G router?

A business 5G router is a purpose-built networking device that combines a 5G cellular modem with routing, security, and management features designed for commercial deployment rather than domestic use. Consumer 5G routers prioritise simplicity and price; business 5G routers prioritise reliability, security, manageability at scale, and durability in demanding environments.

In practice, this means business-grade devices typically include:

• Dual-SIM support with automatic carrier failover, allowing a backup cellular connection if the primary SIM loses coverage or signal
• Multi-WAN failover between 5G, 4G, and fixed-line Ethernet, so no single connectivity path is a single point of failure
• Enterprise VPN support (IPsec, OpenVPN, WireGuard) for secure site-to-site or remote access connectivity
• Cloud-based remote management, enabling configuration, monitoring, firmware updates, and diagnostics across a fleet without requiring physical access to each device
• Industrial-grade hardware ratings for temperature, power input, and physical durability, particularly for vehicles, cabinets, or outdoor enclosures

This is a meaningfully different product category from a consumer hotspot or a CPE router, even where headline 5G speeds might appear similar.

How business 5G routing works

In a typical business 5G deployment, the router connects to the cellular network via one or two SIM cards, negotiates a data session with the operator, and presents a standard IP network to the devices behind it. The router manages WAN selection - preferring 5G where available and falling back to 4G LTE automatically - and handles NAT, firewall, VPN, and DHCP functions for the local network.

Where a fixed-line broadband connection is also present, the router can be configured to use cellular as a primary connection, as a failover backup, or as part of a load-balanced arrangement that spreads traffic across both. For branch sites where fibre is unavailable or prohibitively expensive, 5G as primary WAN is an increasingly practical architecture given current UK coverage levels.

It is important to note that, according to Ofcom's Mobile Matters 2025 research 5G Standalone (SA) downloads are on average 45% faster than 5G Non-Standalone (NSA). Business router selection should take into account whether the device supports SA mode, particularly as operators extend their SA network rollouts through 2026.

Key features and specifications to evaluate

5G standard and band support

• SA (Standalone) and NSA (Non-Standalone) architecture support - SA offers lower latency and better performance where available
• Sub-6 GHz band support for wide-area coverage; mmWave for high-density urban deployments (less relevant for most business use cases in the UK currently)
• Backward compatibility with 4G LTE Cat 20 and 3G for areas where 5G has not yet reached

Dual-SIM and failover

• Dual physical SIM slots with automatic SIM switching and operator black/white listing
• Multi-WAN failover between cellular and Ethernet WAN, with configurable failover triggers (ping check, signal threshold, data limits)
• Band locking capability to pin the modem to specific frequency bands where the operator's 5G coverage is known to be strongest at a given site

Security

• Stateful firewall, IDS/IPS, and application-level filtering for enterprise-grade network perimeter control
• VPN support: IPsec, OpenVPN, WireGuard, and ZeroTier are the most commonly required protocols for business site-to-site and remote access deployments
• Regular firmware updates and a manufacturer with a clear security disclosure process

Remote management

For any deployment involving more than a handful of devices, remote management is essential. The ability to push firmware updates, change configurations, monitor signal quality, and reboot devices remotely makes the difference between a manageable network and one that requires a site visit for every issue. Teltonika's RMS platform and Lantronix's ConsoleFlow each address this need from different perspectives. RMS excels at high-volume IoT and industrial fleet management, while ConsoleFlow offers zero-touch provisioning and SaaS-based device management ideal for enterprise and mission-critical deployments.

Physical and environmental ratings

• Operating temperature range suitable for the deployment environment — a cabinet in an unheated plant room or a vehicle installation can see temperatures well outside the range of a consumer device
• Wide DC voltage input with reverse polarity and overvoltage protection for vehicle or industrial power environments
• Antenna connector type and count (SMA or RP-SMA) and MIMO capability — 4x4 MIMO provides significantly better throughput and reliability than 2x2 in real-world conditions

Benefits of a purpose-built business 5G router

• Reliable primary connectivity at sites where fibre is unavailable, expensive, or subject to long provisioning lead times
• Fast, resilient failover that prevents downtime when a primary fixed-line connection drops
• Secure remote access and site-to-site VPN without depending on a fixed IP address from the operator
• Scalable fleet management without proportional growth in on-site IT overhead
• Future-proofed investment: 5G networks in the UK continue to expand, meaning a device purchased in 2026 will access progressively better coverage and speeds over its operational life

Business 5G router use cases

• Branch and temporary office connectivity: 5G as primary or backup WAN for sites without fibre, pop-up offices, or temporary site offices on construction or events projects
• Vehicle and fleet networking: In-vehicle routers providing crew connectivity, live CCTV uplink, telematics data, and GNSS tracking for transport, logistics, and public safety fleets
• Industrial IoT backhaul: 5G routers aggregating sensor and controller data from a site and transmitting it to a cloud platform or head office SCADA system
• Retail and hospitality: Resilient connectivity for card payment systems, EPOS, digital signage, and guest Wi-Fi where a fixed-line outage would directly impact trading
• Critical infrastructure and utilities: Remote sites such as substations, pumping stations, and renewables installations where cellular backhaul is the only practical connectivity option

5G Business router comparison: Options from BCD Atlantik

The Teltonika RUTX50 is the most versatile option for deployments requiring strong industrial credentials alongside 5G, particularly where a highly configurable open platform and RMS remote management matter. The TRB500 suits scenarios where a compact 5G uplink is needed behind an existing router or connected directly to a single device. The Lantronix G520 is the stronger choice for Industry 4.0 and critical infrastructure applications, where built-in cybersecurity features, fieldbus protocol conversion, and ConsoleFlow's zero-touch provisioning reduce integration effort and ongoing management overhead. The E224 complements this for enterprise branch deployments where serial device connectivity alongside cellular WAN is a requirement.

A note on UK 5G coverage in 2026

When specifying a 5G router for a fixed site, it is worth checking coverage from multiple operators before committing to a SIM contract. Ofcom estimates that only 62% of the UK's land mass currently has 5G service available from at least one provider, which means rural and semi-rural sites may find 4G LTE remains the primary cellular technology available. A dual-SIM router with strong 4G fallback is therefore still the right specification for many UK deployments, even where 5G is the target technology.

Following the completion of the Vodafone and Three merger in mid-2025, the UK mobile market has entered a period of structural change, with network integration still under way. The resulting VodafoneThree network is the largest in the UK by subscribers and is expected to accelerate 5G rollout through combined spectrum and infrastructure. This makes dual-SIM capability with multi-operator support a sensible investment for deployments intended to remain in service for three to five years.

5G Business router buying guide: Key questions to answer

1. Fixed site or mobile? Vehicle and mobile deployments need a device rated for vibration, wider temperature ranges, and vehicle power inputs. Fixed-site deployments prioritise Ethernet port count and mounting options.
2. How many devices do you need to manage? For more than five or six routers, cloud-based remote management (Teltonika RMS or Lantronix ConsoleFlow) becomes essential rather than optional. ConsoleFlow is available as both SaaS and on-premises deployment for organisations with data sovereignty requirements.
3. Do you need Wi-Fi at the router, or will existing Wi-Fi infrastructure handle local wireless? The TRB500 is a pure 5G uplink gateway with no Wi-Fi; the RUTX50 includes dual-band Wi-Fi; the Lantronix G520 includes Wi-Fi depending on variant.
4. Do you need fieldbus protocol support or serial device integration? The Lantronix G520's fieldbus conversion suite and the E224's serial connectivity (RS232/RS485) are relevant for sites integrating legacy industrial equipment alongside cellular WAN.
5. Which operators have 5G at your target sites? Confirm this before selecting SIMs; the router hardware is only part of the equation.
6. What is the total cost of ownership? Factor in hardware, management platform subscriptions, SIM contracts, and installation when comparing options; headline hardware prices rarely tell the full story.

Frequently asked questions

What is the difference between a 5G business router and a 5G consumer router?

Business routers include dual-SIM failover, enterprise VPN, remote fleet management, more Ethernet ports, and hardware rated for industrial or vehicle environments. Consumer devices are designed for home use, typically with a single SIM, basic Wi-Fi, and no remote management capability.

Do I need a 5G router if my site already has fibre broadband?

Not necessarily as a primary connection, but a 5G router as a cellular failover can prevent costly downtime if your fixed-line circuit goes down. For businesses where connectivity outages directly impact trading or operations, the cost of a 5G failover device is typically low relative to the cost of an outage.

What is the difference between 5G SA and 5G NSA?

5G Non-Standalone (NSA) uses the 4G LTE network for control signalling and adds 5G for data throughput. 5G Standalone (SA) is a native 5G network architecture that delivers lower latency and higher efficiency. Ofcom data shows SA downloads are around 45% faster than NSA on average, and SA is the architecture being expanded by UK operators through 2026.

Can a 5G router replace a fibre broadband connection for business?

In many cases, yes — particularly for sites where fibre is unavailable or subject to long lead times. 5G download speeds are now sufficient for most business workloads including VoIP, video conferencing, and cloud applications. Latency and upload speeds are the main considerations; some high-demand workloads may still favour a wired connection where one is available.

Is Teltonika RMS included with the RUTX50?

The RMS platform is a subscription service from Teltonika, separate from the router hardware. It provides cloud-based remote management for device configuration, monitoring, and firmware updates. Contact BCD Atlantik for current RMS licensing and pricing details, as well as RUTX50 hardware quotations.

What is Lantronix ConsoleFlow and does the G520 require a subscription?

ConsoleFlow is Lantronix's cloud device management and connectivity services platform, available as both a SaaS application and an on-premises deployment. The G520 comes pre-enabled for ConsoleFlow, providing zero-touch provisioning, remote configuration, firmware updates, and device telemetry from day one. It suits organisations that need centralised visibility across distributed sites without building custom management infrastructure. Contact BCD Atlantik for current ConsoleFlow licensing details alongside G520 hardware pricing.

Which UK networks provide the best 5G coverage for business in 2026?

Based on Ofcom's late 2025 figures, EE leads for outdoor premises coverage at 89%, followed by O2 at 76%. For any specific site, checking coverage directly with operators and ideally conducting a site survey with a test SIM is the most reliable approach before committing to a 5G deployment.

Can the Teltonika TRB500 be used as a 5G modem behind an existing router?

Yes. The TRB500's single Gigabit Ethernet port can connect to an existing router's WAN input, providing a 5G uplink without replacing the downstream network. This makes it a practical option for adding 5G capability to an established network without a full infrastructure change.

Key takeaways

• 5G coverage in the UK has expanded materially through 2025 and 2026, making a 5G router for business a practical choice for primary or failover WAN across a much wider range of sites than a few years ago
• Dual-SIM with automatic 4G failback remains essential, as 5G coverage is still unavailable across significant portions of the UK's geographic landmass
• The Teltonika RUTX50 offers the strongest combination of 5G capability, interface flexibility, and open configurability for industrial and IoT-oriented deployments
• The Lantronix G520 suits Industry 4.0, critical infrastructure, and transportation deployments where built-in security, fieldbus protocol conversion, and ConsoleFlow cloud management are priorities
• Total cost of ownership should include router hardware, any management platform subscription, SIM contracts, and installation; hardware price alone is not a complete comparison
• 5G Standalone (SA) mode delivers materially better performance than NSA where available; confirm which architecture your target operator supports at your specific sites

Conclusion

A business 5G router is now a realistic part of most UK network infrastructure decisions, whether as a primary WAN connection, a resilient failover, or a connectivity backbone for remote and mobile deployments. The right choice depends less on headline 5G speeds and more on the deployment environment, management requirements, security posture, and the specific coverage available from your chosen operators at the sites you need to connect.

BCD Atlantik supplies 5G routers and gateways from Teltonika and Lantronix, available for B2B purchase with project support from our team. Browse our full 5G and cellular router range, or contact us to discuss the right specification for your deployment.

Raspberry Pi Compute Module business use has grown significantly over the past few years, moving well beyond education and hobbyist prototyping into serious commercial product development. Manufacturers, system integrators, and engineering teams across sectors from healthcare to industrial automation are now designing the Raspberry Pi Compute Module into products they intend to ship in volume. The reason is straightforward: the Compute Module takes the processing power, software ecosystem, and community support of the Raspberry Pi platform and repackages it into a compact, integration-ready form that suits custom hardware design rather than off-the-shelf development boards.

This guide explains what the Raspberry Pi Compute Module is, why it has become a credible choice for embedded systems in a business context, and how to decide which variant fits your application.

Summarise Raspberry Pi Compute Module insights in ChatGPT

Summary

The Raspberry Pi Compute Module is a system-on-module (SOM) that combines the processor, RAM, and optional eMMC storage of a Raspberry Pi onto a compact board designed to plug into a custom carrier board rather than be used standalone. For businesses, it offers a well-supported, cost-effective platform for building embedded products in volume, with the advantage of a mature Linux software stack, long-term supply commitments, and a large developer community. The CM4 and CM5 are the two current generations, differing primarily in processing power, I/O capability, and target application profile.

What is a Raspberry Pi Compute Module?

A Raspberry Pi Compute Module is a small form-factor system-on-module that packages the core components of a Raspberry Pi (processor, memory, and storage) onto a board designed for integration rather than direct use. Unlike the standard Raspberry Pi single-board computers, a Compute Module has no USB ports, no full-size HDMI output, and no GPIO headers in the traditional sense. Instead, it exposes its interfaces through a high-density connector, allowing an engineer to design a custom carrier board that breaks out only the interfaces the product actually needs.

This distinction matters commercially. A standard Raspberry Pi is designed as a general-purpose development board; a Compute Module is designed to become a component inside something else. That shift in intent changes what becomes possible in terms of product form factor, interface selection, environmental ruggedisation, and BOM optimisation.

How it works in an embedded product

In a typical embedded deployment, the Compute Module sits on a custom-designed carrier board developed by the product manufacturer. The carrier board provides the connectors, power management, peripherals, and industrial interfaces relevant to that specific product; whether that is a touchscreen interface, RS485 for machine communication, a cellular modem, or an array of digital I/O lines for control applications.

During development, teams commonly use an official IO board, such as the Raspberry Pi CM4 IO Board or the Raspberry Pi CM5 IO Board, to prototype and validate their software before committing to a custom PCB design. Once the software stack is stable and the hardware requirements are confirmed, the carrier board can be designed to exactly match the product's needs, removing any unnecessary components and reducing unit cost accordingly.

The Raspberry Pi OS and its ecosystem of libraries, drivers, and community tooling transfer directly from the development environment to production, significantly reducing the software bring-up time compared to platforms with less established support.

Key features and specifications

Raspberry Pi CM4

The Compute Module 4 is built around a quad-core ARM Cortex-A72 processor running at 1.5GHz, with up to 8GB of LPDDR4 RAM and optional eMMC storage in 8GB, 16GB, or 32GB capacities. It supports dual 4K display output, hardware video decode at up to 4Kp60, Gigabit Ethernet, USB 2.0, dual camera interfaces, and a PCIe Gen 2 x1 interface for expansion. Wireless variants include dual-band 2.4/5GHz Wi-Fi and Bluetooth 5.0 with modular compliance certification, reducing the regulatory overhead of adding wireless connectivity to a product.

Available in a number of variants covering different combinations of RAM, storage, and wireless, the CM4 gives engineering teams flexibility to match the module specification tightly to the application, avoiding the need to pay for memory or storage that the product does not require. Extended temperature variants rated from -40°C to +85°C are also available through BCD Atlantik, making the CM4 suitable for demanding environmental conditions.

Raspberry Pi CM5

The Raspberry Pi Compute Module 5 represents a significant step up in performance. Built on the same quad-core Arm Cortex-A76 architecture as the Raspberry Pi 5, it delivers noticeably higher compute throughput and improved I/O capabilities compared to the CM4. RAM options extend to 16GB, and eMMC storage is available from 0GB Lite up to 64GB. Dual 4Kp60 HDMI output, Gigabit Ethernet, and the option of a fully-certified wireless module round out the specification.

The CM5 is the right choice where the application needs more compute headroom, such as handling multiple data streams, running edge inference workloads, or driving high-resolution display content alongside background processing tasks. For teams starting a new design today, the CM5 offers a longer effective product life before compute becomes a constraint.

Benefits of using Raspberry Pi Compute Modules for business applications

• Established ecosystem: The Raspberry Pi software stack, community, and library support are among the most mature of any embedded Linux platform, reducing development risk and recruitment difficulty.
• Flexible variant selection: Specifying the right combination of RAM, storage, and wireless avoids over-engineering the BOM for applications with modest requirements.
• Modular compliance certification: Wireless-equipped variants carry pre-certified RF compliance, reducing the regulatory testing burden when adding Wi-Fi or Bluetooth to a product.
• Long-term supply commitment: Raspberry Pi has committed to long-term availability of its Compute Modules, which matters considerably for embedded products with multi-year production runs.
• Separation of hardware and software development: Software can be developed and validated on an IO board while carrier board hardware is being designed in parallel, compressing the overall development timeline.
• Carrier board design freedom: Exposing all interfaces through a single high-density connector means the carrier can be optimised entirely around the product's requirements, with no compromises inherited from a general-purpose board layout.

Applications and use cases

Raspberry Pi CM business applications span a broad range of industries and product types. Some of the most common include:

• Industrial HMI and control panels: The CM5-based Edatec ED-HMI3120 series integrates the Compute Module into a ruggedised touchscreen display with industrial power input, watchdog, and crypto authentication - a ready-made embedded HMI solution for machine builders
• Edge gateways and connectivity computers: The Edatec ED-IPC1100 and ED-IPC3100 series use the CM0 and CM5 respectively, combining fieldbus interfaces, 4G connectivity, and DIN-rail mounting to create deployable industrial computers based on the Raspberry Pi platform.
• Medical and diagnostic devices: The combination of a controlled software environment, modular compliance certification, and established Linux drivers makes the CM platform attractive for Class I medical device integration.
• Digital signage and kiosk systems: Dual 4K display output and reliable long-term supply make the CM4 and CM5 well-suited to commercial display products requiring multi-year consistency.
• Retail and point-of-sale terminals: Low power draw, a compact footprint, and strong USB peripheral support suit embedded POS applications where the hardware must disappear inside a polished enclosure.
• Building automation and smart infrastructure: Wireless connectivity, Ethernet, and broad GPIO availability through a carrier board make the Compute Module a practical foundation for building management systems and access control devices.

CM4 vs CM5: Which is right for your project?

Feature	Raspberry Pi CM4	Raspberry Pi CM5
Processor	Quad-core ARM Cortex-A72, 1.5GHz	Quad-core Arm Cortex-A76, 2.4GHz
RAM options	1GB, 2GB, 4GB, 8GB	2GB, 4GB, 8GB, 16GB
eMMC storage	0GB Lite, 8GB, 16GB, 32GB	0GB Lite, 8GB, 16GB, 32GB, 64GB
Display output	Dual 4K	Dual 4Kp60 HDMI
PCIe	Gen 2 x1	Gen 3 x1 (via IO board)
Wireless (optional)	Wi-Fi 5 + BT 5.0	Wi-Fi 5 + BT 5.0
Extended temp option	Yes, -40°C to +85°C (contact BCD Atlantik)	Not yet available
Best suited for	Cost-sensitive, stable designs; extended temp applications	Higher compute demand; new designs with longer roadmap

For most new embedded product designs starting today, the CM5 is the stronger foundation — particularly where the product roadmap extends several years and compute requirements may grow. The CM4 remains highly relevant for cost-sensitive applications, designs already in production, and projects requiring the extended temperature range variants.

Getting Started: Development Kit and IO boards

For teams new to the Compute Module platform, or moving from CM4 to CM5, the Raspberry Pi CM5 Development Kit provides a complete starting point: a CM5 module and IO board with all the accessories needed to begin software development and hardware prototyping immediately. It is the most efficient way to validate the software stack before investing in carrier board design.

The CM4 IO Board and CM5 IO Board both expose the full interface set of their respective modules, providing a reference design that carrier board engineers can use to understand interface routing, power requirements, and recommended component selection before committing to a custom layout.

Best practices for Raspberry Pi CM business deployments

1. Define your interface requirements before selecting a variant. The carrier board design is driven by your I/O needs; the CM selection should follow from that, not the other way around.
2. Choose the Lite (0GB eMMC) variant if using external storage. Where the product will boot from an NVMe SSD or SD card via the carrier board, the Lite variant avoids paying for eMMC that will not be used.
3. Account for thermals early. Compute Modules in enclosed carrier board designs may need active or passive thermal management; the development kit phase is the right time to characterise this, not after tooling the enclosure.
4. Validate modular compliance early if shipping wireless products. The certified wireless module simplifies regulatory approval, but you still need to confirm the antenna design and placement on your carrier board do not invalidate the certification.
5. Plan for long-term software maintenance. Choose an OS distribution with a clear long-term support commitment; the Raspberry Pi OS LTS releases or Ubuntu LTS are both well-supported choices for production embedded deployments.
6. Register your project with your distributor. BCD Atlantik's project registration process for B2B CM customers ensures priority access to stock and supports volume pricing discussions before production ramp.

Frequently asked questions

What is the difference between the Raspberry Pi CM4 and a standard Raspberry Pi 4?

The CM4 uses the same processor and core silicon as the Raspberry Pi 4 Model B but removes the standard connectors and packages everything into a compact module designed to plug into a custom carrier board. It is not meant to be used standalone; it is a component for integration into a product.

Is the Raspberry Pi Compute Module suitable for commercial production runs?

Yes. As well as individually boxed modules, you can also purchase compute modules in bulk packs, where they come in trays. This reduces the time and cost of opening individual retail packages, and means they can easily be incorporated into a production workflow.

In addition, Raspberry Pi has committed to long-term supply of its Compute Modules specifically to support commercial and industrial customers with multi-year production needs. BCD Atlantik can support volume & bulk pack procurement and project registration for B2B customers.

Can I upgrade from CM4 to CM5 without redesigning my carrier board?

The CM5 uses a different connector than the CM4 and is not a drop-in replacement. A carrier board redesign is required. However, the software migration path is straightforward given the shared Raspberry Pi OS ecosystem.

What operating system do Raspberry Pi Compute Modules run?

Raspberry Pi OS (based on Debian) is the primary supported OS and is well-suited to embedded Linux deployments. Ubuntu and other Linux distributions are also supported on the CM platform.

Are extended temperature variants of the CM4 available?

Yes. Since March 2025, extended temperature CM4 variants rated from -40°C to +85°C have been available. These are not listed as standard shop variants; contact BCD Atlantik directly for pricing and availability.

What is the difference between the eMMC and Lite (0GB) variants?

The eMMC variants include onboard flash storage for the OS and application data. The Lite variant omits this, requiring the carrier board to provide an alternative boot storage, typically via an SD card slot or NVMe SSD on the carrier. Lite variants cost less and are preferred where external storage is already part of the design.

Do I need a development kit to get started?

Not strictly, but it is strongly recommended. The CM5 Development Kit provides the IO board, module, and accessories needed to begin software development and hardware bring-up without designing a carrier board first. It significantly reduces the time from evaluation to a validated software stack.

Can BCD Atlantik support ODM and custom hardware development around the Compute Module?

Yes. Alongside distribution, BCD Atlantik supports custom hardware development projects, including concept design, prototype development, and production support for products built on the Raspberry Pi Compute Module platform.

Key takeaways

• Raspberry Pi Compute Modules are designed for integration into custom carrier boards, not standalone use - this makes it well-suited to embedded product development.
• The CM4 remains a strong choice for cost-sensitive, extended-temperature, or already-in-production designs; the CM5 is the right foundation for new designs requiring more compute headroom.
• Modular wireless compliance certification, a mature Linux software stack, and long-term supply commitments are key commercial advantages of the Compute Module platform.
• Development kits and IO boards allow software development to proceed before carrier board hardware is finalised, compressing overall project timelines.
• Raspberry Pi CM business applications span HMI, edge computing, medical devices, digital signage, and industrial connectivity - with ready-made carrier solutions such as the Edatec IPC and HMI series also available

Conclusion

The Raspberry Pi Compute Module has earned its place in commercial embedded system design by combining genuine processing capability with an accessible software ecosystem, a flexible variant structure, and supply chain commitments that make it credible for multi-year production. Whether you are building an industrial HMI, an edge gateway, a connected kiosk, or a custom embedded controller, the CM platform offers a faster route from prototype to production than many proprietary alternatives.

BCD Atlantik is an authorised Raspberry Pi reseller supplying the full Compute Module 4 and Compute Module 5 range, including Lite and wireless variants, IO boards, development kits, and extended temperature CM4 options. We also stock the Edatec range of CM5-based industrial computers and HMI displays for teams looking for a ready-made carrier solution. Browse our full Raspberry Pi embedded range or get in touch to discuss your project requirements.

Ask AI to explain Raspberry Pi Compute Modules

An IoT gateway for industrial use sits at the boundary between your machinery and your data systems, translating signals from sensors, PLCs and legacy equipment into a format your monitoring platform or cloud service can actually use. Choosing the wrong one is rarely obvious at first; it tends to show up months later as dropped connections, security gaps, or a device that simply cannot keep up as the deployment grows.

This guide walks through what an industrial IoT gateway actually does, the specifications that matter most, and how to evaluate options against your own application before you commit budget to hardware that will likely be in the field for five to ten years.

Summarise IoT gateway insights in ChatGPT

Summary

The right industrial IoT gateway for your application depends on five factors: the connectivity protocols your equipment already speaks (Modbus, CAN, RS485, etc.), the environmental conditions it will face (temperature, vibration, ingress), the network options you need (Ethernet, Wi-Fi, 4G/5G), the level of edge processing required, and the security and certification standards your industry demands. For most industrial sites, that means a DIN-rail mountable gateway rated for at least -20°C to 60°C, with wide-voltage DC input, hardware-level security, and remote management support rather than a consumer-grade router repurposed for the job.

What is an industrial IoT gateway?

An industrial IoT gateway is a ruggedised device that connects field-level equipment, such as sensors, actuators, PLCs and meters, to local networks, cloud platforms, or both. Unlike a standard network router, it is built to operate continuously in environments with extreme temperatures, vibration, electrical noise, and limited or no on-site IT support.

Functionally, an industrial gateway typically performs three jobs at once:

• Protocol translation, converting industrial fieldbus protocols such as Modbus RTU/TCP, CAN, or RS232/RS485 into IP-based formats like MQTT or HTTP

• Connectivity, providing a reliable uplink via Ethernet, Wi-Fi, or cellular (4G/5G) with failover between them
• Edge processing, filtering, aggregating, or acting on data locally before it reaches the cloud, reducing bandwidth use and latency

This is different from a generic IoT router, which usually just provides connectivity, or an industrial PC, which offers full compute but at higher cost and power draw. A gateway sits between the two: enough processing for protocol handling and light analytics, with the durability and interfacing that industrial environments demand.

How industrial IoT gateways work

In a typical deployment, sensors and controllers on the plant floor connect to the gateway over a wired fieldbus or local wireless link. The gateway normalises that data, applies any local logic (threshold alerts, basic filtering, store-and-forward buffering during connectivity drops) and forwards it upstream to a SCADA system, historian, or cloud IoT platform.

For example, a water treatment facility might use a gateway to poll a dozen Modbus RTU flow meters every few seconds, buffer the readings locally if the cellular connection drops, and push aggregated data to a cloud dashboard once connectivity is restored. The gateway is doing the unglamorous but essential work of keeping data flowing reliably, even when the network around it is not.

Key features and specifications to evaluate

When comparing industrial IoT gateways, the following specifications tend to separate genuinely industrial-grade hardware from consumer products in industrial packaging.

Connectivity and protocol support

• Wired interfaces: Ethernet (1x or more, ideally Gigabit), RS232/RS485, CAN
• Wireless: Wi-Fi, Bluetooth, 4G LTE Cat 1/4/6 or 5G depending on bandwidth needs
• Protocol support: Modbus RTU/TCP, MQTT, SNMP, and increasingly OPC-UA for newer automation systems
• Dual-SIM with automatic failover, useful for sites without a fixed-line backup

Environmental and mechanical durability

• Operating temperature range of at least -20°C to 60°C, with -40°C to 70°C for extreme climates or outdoor installations
• Wide DC input voltage (commonly 9–36V) with reverse polarity, overvoltage and overcurrent protection
• All-metal enclosure with DIN-rail mounting for cabinet installation
• IP-rated ingress protection where the gateway will be exposed to dust or moisture

Security

• Hardware-based crypto authentication and secure boot
• VPN support (IPsec, OpenVPN, WireGuard) for secure remote access
• Firewall and access control, with regular firmware updates from the manufacturer
• Remote management platform support for fleet-wide configuration and monitoring

Edge compute and AI capability

Increasingly, industrial gateways are expected to do more than relay data. Where a deployment involves machine vision, predictive maintenance, or anomaly detection, the gateway needs enough onboard processing to run inference locally rather than sending every frame to the cloud. This is where edge AI-capable platforms, built on processors such as the Qualcomm Dragonwing IQ series, are becoming increasingly relevant for gateway-class devices rather than just industrial PCs.

Benefits of choosing the right gateway

• Reduced downtime, through reliable failover connectivity and local data buffering during outages
• Lower long-term cost, by avoiding premature hardware failure or replacement in harsh conditions
• Easier scaling, with remote management reducing the need for site visits as the deployment grows
• Stronger security posture, limiting exposure as more OT devices connect to IT networks
• Faster decision-making, where edge processing reduces the latency between an event and a response

Applications and use cases

Industrial IoT gateways are used across a wide range of sectors, including:

• Manufacturing: Connecting legacy PLCs and sensors to modern MES or SCADA systems without replacing existing equipment
• Energy and utilities: Remote monitoring of substations, pumping stations and renewable assets in locations with limited connectivity
• Transportation and logistics: Fleet tracking, cold chain monitoring, and remote diagnostics on vehicles and containers
• Building and infrastructure management: HVAC, lighting and access control integration across distributed sites
• Agriculture: Soil and environmental sensor networks in remote, low-connectivity locations

Comparison: Gateway options from BCD Atlantik

To make this concrete, here is how three different gateway classes available from BCD Atlantik compare against typical industrial requirements.

The Teltonika TRB and RUT series suit applications that are primarily about reliable connectivity and remote management. The Edatec ED-IPC range adds more onboard compute and isolated I/O for sites running fieldbus protocols alongside connectivity. Where AI inference at the edge is part of the requirement, Dragonwing IQ-based platforms extend that further with substantial on-device AI performance and industrial temperature tolerance, while remaining well short of full industrial PC cost and power draw.

Best practices for selecting an industrial IoT gateway

1. Map your existing protocols first. List every fieldbus or interface your equipment already uses before evaluating hardware; retrofitting protocol support later is expensive.
2. Plan for the environment, not the spec sheet. If summer ambient temperatures inside a cabinet regularly exceed 50°C, build in headroom rather than choosing a gateway rated right at that limit.
3. Prioritise remote manageability. A device fleet without remote configuration and firmware update support becomes a maintenance liability as it scales beyond a handful of sites.
4. Check long-term availability. Industrial deployments often run for a decade; choose manufacturers with clear product longevity commitments rather than consumer-cycle hardware.
5. Don't over-spec edge compute you don't need.Edge AI capability adds cost and complexity, only worth it where local inference genuinely reduces latency or bandwidth requirements.
6. Confirm certifications relevant to your sector, such as CE, FCC, ATEX, or industry-specific standards before procurement, not after.

Frequently asked questions

What is the difference between an IoT gateway and an industrial router?

A router primarily provides network connectivity. A gateway typically adds protocol translation and edge processing on top of connectivity, making it better suited to integrating legacy industrial equipment with modern data platforms.

Can an industrial IoT gateway work without an internet connection?

Yes. Most industrial gateways support local data buffering and store-and-forward functionality, so data collection continues during connectivity outages and resumes uploading once the connection is restored.

What temperature range should I look for in an industrial gateway?

For most indoor industrial environments, -20°C to 60°C is a reasonable baseline. Outdoor, vehicle-mounted, or extreme-climate applications typically need -40°C to 70°C or wider.

Do industrial IoT gateways support 5G?

Increasingly, yes. 5G-capable gateways such as the Teltonika TRB500 and RUTX50 are available where high bandwidth or low latency is required, though many industrial applications are still well served by 4G LTE.

How important is cybersecurity in gateway selection?

Very. As OT and IT networks converge, gateways are a common attack surface. Look for hardware-based authentication, VPN support, regular firmware updates, and a manufacturer with a clear security disclosure process.

Can one gateway support multiple protocols at once?

Most industrial gateways support several protocols simultaneously, such as Modbus RTU over RS485 alongside MQTT over cellular, though the exact combination depends on the hardware and firmware.

What is edge AI doing in a gateway, and do I need it?

Edge AI allows the gateway to run inference, such as defect detection or anomaly recognition, locally rather than sending raw data to the cloud. It is worth the added cost where latency, bandwidth, or data privacy make cloud-only processing impractical; for straightforward monitoring and control, it is usually unnecessary.

How long do industrial IoT gateways typically last in deployment?

Well-specified industrial gateways are commonly deployed for five to ten years, particularly where the manufacturer commits to long-term product availability and firmware support.

Key takeaways

• An industrial IoT gateway combines protocol translation, connectivity and edge processing ‐ it is not simply a ruggedised router
• Match connectivity and protocol support to your existing field equipment before evaluating other specifications
• Environmental rating, power input flexibility and DIN-rail mounting are non-negotiable for most industrial sites
• Security and remote manageability become more important as deployments scale beyond a single site
• Edge AI capability is increasingly available in gateway-class hardware, but should be matched to genuine application need rather than added by default

Conclusion

Choosing the right IoT gateway for an industrial application comes down to matching hardware to the realities of your site: the protocols already in use, the environmental conditions, the connectivity options available, and how much processing needs to happen at the edge rather than in the cloud. Getting this right at the evaluation stage avoids costly retrofits later and sets up a deployment that can scale reliably over its full operating life.

BCD Atlantik supplies industrial IoT gateways and connectivity computers from Teltonika and Edatec, alongside Qualcomm Dragonwing-based edge compute platforms for applications requiring on-device AI. If you're evaluating options for a specific site or application, our team can help match the right hardware to your requirements.

Ask AI to explain industrial IoT gateways