Ezurio Launches BL54LM20 Series with More Memory and On-Device Processing for IoT

Semicon Leaders Asia speaks with Jonathan Kaye, VP of Product Management and Marketing at Ezurio, about the company’s BL54LM20 Series, growing demand for higher memory and connectivity in IoT devices, the move towards on-device processing, and the opportunities for secure, low-power Edge AI across connected products.

Q. What customer and market needs are driving the launch of the BL54LM20 Series, and how does the combination of connectivity, additional memory and Edge AI address the changing requirements of IoT product developers?

Since the launch of our Nordic nRF54L family of modules - BL54L10 / L15 / L15µ, our customers have told us the same thing: their applications are outgrowing the memory and I/O headroom of a typical Bluetooth LE module long before they outgrow the wireless performance itself. Firmware footprints are getting larger because products now need to support multiple protocols, richer over-the-air update mechanisms, more sophisticated security stacks and increasingly complex application logic, all on the same certified module they started with. At the same time, a growing number of those customers are asking a second question: can we start doing some of our sensor processing or classification locally, on the device, rather than shipping raw data to the cloud.

The BL54LM20 Series was built to answer both of those needs in a single family. The BL54LM20A gives developers 2MB of non-volatile memory and 512KB of RAM, a 128 MHz Arm Cortex-M33 application processor paired with a 128 MHz RISC-V coprocessor, USB, and an expanded I/O set. That is a substantial step up in headroom for larger applications, richer protocol combinations such as Bluetooth LE, Thread, Matter and NFC running together, and for gateway-class or controller-style designs that previously would have needed to move to a heavier compute platform.

The BL54LM20B takes the same silicon and hardware platform and adds Nordic's Axon NPU, which is where the Edge AI piece comes in. Rather than asking customers to bolt on a separate AI processor or offload inference to the cloud, we are putting real machine learning capability directly onto the connectivity module they already trust for RF performance, certification and security. That combination, more memory and I/O for the applications customers are building today, plus a clear, low-risk path to on-device AI for what they want to build next, is exactly what the market is asking for. It lets a product team solve today's memory and integration problem without betting the design on an experimental AI chipset, because the AI capability lives on the same pre-certified, footprint-compatible module family.

Q. The BL54LM20A and BL54LM20B are footprint-compatible and share the same RF design and certification. How does this approach help customers start with a lower-complexity design today while keeping a clear path towards Edge AI in the future?

This is one of the most important decisions we made in defining this series, and it goes directly to how we think about de-risking product development for our customers. Certification is expensive and slow. RF requalification, regulatory testing across regions, Bluetooth SIG qualification- all of that consumes budget and calendar time that most IoT product teams simply do not have to spare. Historically, if a customer wanted to add capability that was not on their original module, they faced board respins and, in many cases, a new certification cycle.

Because the BL54LM20A and BL54LM20B are 100% footprint-compatible and share the same RF design and certification, that problem goes away. A customer can lay out their board once, launch with the BL54LM20A to get the memory, I/O and connectivity improvements they need immediately, and know that if their product roadmap calls for on-device AI in twelve or eighteen months, they can move to the BL54LM20B without touching the PCB layout and without restarting the certification clock. The RF performance, the antenna options, the regulatory approvals, all of it carries forward.

That matters commercially as well as technically. It means a customer's initial bill of materials decision is not really a bet on whether they will need Edge AI. They do not have to over-specify the module today just in case, and they do not have to under-specify it and risk a costly redesign later. They can sequence their investment to match their actual product roadmap. For product managers and engineering leads under pressure to hit a launch date, that is a meaningful reduction in schedule risk and NRE exposure, and it is a big part of what we mean when we talk about the BL54LM20 Series lowering the barrier to Edge AI for every Bluetooth LE developer, not just the teams with dedicated AI engineering resources.

Q. The BL54LM20B brings Nordic's Axon NPU into a pre-certified module and offers up to 15x faster on-device inference compared with CPU processing alone. Which real-world applications do you expect to benefit most from this capability?

We see the strongest early pull from four areas. Predictive maintenance sensors are probably the clearest fit. These devices are typically battery-powered, need to run continuously, and are most valuable when they can flag an anomaly, a bearing vibration signature or a temperature trend the moment it happens rather than after a batch upload to the cloud. Faster on-device inferencing means the sensor can classify what it is seeing in real time, at a fraction of the power cost of running that same workload on a general-purpose CPU.

Wearable and portable medical devices are the second area. These products are held to a high bar on responsiveness, privacy and battery life all at once, and Edge AI on a pre-certified module lets a medical device manufacturer keep patient data on the device, respond to a physiological event immediately, and yet keep within a constrained power budget.

Factory automation and robotics are the third. Anywhere a machine needs to make a local decision, detecting a part misalignment, recognising a gesture or a fault condition, latency is the enemy, and round-tripping to the cloud simply is not fast enough or reliable enough on a factory floor. The fourth is smart building and commercial IoT, where occupancy sensing, acoustic event detection and similar classification tasks benefit from staying local both for responsiveness and for the practical reason that many commercial buildings have inconsistent or congested wireless backhaul.

Across all four, the common thread is the same: developers get up to 15x faster inference than CPU-only processing, without needing to become NPU experts, and without giving up the certified Bluetooth LE, Thread, Matter and NFC connectivity that got them to market in the first place.

Q. For developers working on predictive maintenance, robotics, wearables and smart buildings, what advantages does local AI processing provide in terms of responsiveness, privacy, connectivity requirements and power efficiency?

Each of those four dimensions matters differently depending on the application, but they all point in the same direction, toward keeping intelligence at the edge. On responsiveness, when inference happens on the module itself rather than after a round trip to a cloud service, you eliminate network latency and jitter. For a robotics application reacting to a sensor event, or a wearable detecting a fall or an arrhythmia, milliseconds matter, and local processing removes the uncertainty of a wireless link and a cloud endpoint from the critical path.

On privacy, local inference means raw sensor data, whether that is audio, biometric signals or image-derived features, never has to leave the device to produce a usable result. That is increasingly important for medical wearables and for any smart building deployment where occupants have legitimate concerns about what is being captured and where it is going. It also simplifies a customer's regulatory and data-handling story considerably.

On connectivity requirements, local AI processing means the product does not need to depend on continuous, high-bandwidth connectivity to function. It can operate fully offline, and only needs to communicate over Bluetooth LE, Thread or Matter when it actually has something worth reporting: an alert, a classification result, a status update. That is a much lighter connectivity burden, and it is a real advantage in industrial and building environments where wireless coverage is inconsistent.

Finally, on power efficiency, this is where the Axon NPU earns its keep. Running an inference workload on a dedicated, purpose-built accelerator is dramatically more power efficient than running the same workload on a general-purpose CPU core, and that efficiency gain compounds over the life of a battery-powered device. For always-on machine learning tasks, that is the difference between a sensor that needs a battery change every few months and one that runs for years on the same cell.

Q. The series combines Bluetooth LE, Thread/Matter, NFC, expanded I/O and increased memory in a compact module. How does this broader feature set help customers reduce design complexity and bring connected products to market faster?

The core idea here is flexibility. The BL54LM20 module family gives customers Bluetooth LE Core 6.0, 802.15.4 for Thread and Matter, and NFC, in an 11 by 8 by 1.8mm footprint, so they are not locked into one protocol at the design stage. A customer building a simple Bluetooth LE product uses exactly that, and a customer who needs Thread, Matter or NFC alongside it has that capability already sitting on the same module, without a different part number, a different footprint or a new certification to plan around. That flexibility means the protocol mix can be a software and configuration decision rather than a hardware one, which matters most for teams who are not yet certain which combination their product, or the next revision of it, will actually need.

The memory increase matters just as much as the protocol consolidation. 2MB of non-volatile memory and 512KB of RAM give a development team enough room to run a Matter stack alongside their own application logic, their over-the-air update mechanism and their security services, without constantly fighting a memory budget. That headroom is what allows richer protocol combinations to coexist on one chip in the first place. Combine that with an expanded I/O set, high-speed USB, and a broader set of low-leakage peripherals for sensors and HMIs, and a designer can support more of their product's requirements directly from this one module rather than reaching for an external microcontroller.

The net effect for a customer's time to market is straightforward. Fewer components to qualify, fewer RF interactions to characterise, fewer separate certification submissions to manage, and more of their target feature set already accounted for in a pre-certified, PSA Certified level 3 platform with Secure Boot, Secure Firmware Update and Secure Storage built in. That is design complexity removed before the customer even opens their schematic tool, which translates directly into a shorter path from concept to a certified, shippable product.

Q. With the BL54LM20B USB-C Adapter also providing a pre-certified path for development and production use, what opportunities do you see for Ezurio as more customers look to add secure, low-power Edge AI capabilities to existing connected products?

The USB-C Adapter is a deliberate answer to a practical problem we kept hearing about: customers with an existing PC, gateway or industrial system who want to add Bluetooth LE, Thread, Matter or Edge AI capability to that product without a full board redesign. By embedding the certified BL54LM20B module in a compact, regulatory-approved USB-C form factor, we give those customers a plug-in path to Edge AI that carries no additional certification risk and very little integration effort, whether they are using it for early-stage development, for internal test and production rigs, or as a final shipped accessory for their own product.

That opens up a broader opportunity for Ezurio than the module business alone. Every one of our existing BL54L customers, and a large number of gateway, PC accessory and industrial system vendors who were never going to redesign a board just to add Bluetooth or Edge AI, is now a realistic prospect for a plug-in adapter with the same pre-certified guarantees they already associate with our module line. It lowers the barrier to entry for customers evaluating Edge AI for the first time, because they can prototype and even ship an initial version of a product using the adapter before committing engineering resources to a fully embedded design.

Strategically, this reflects how we think about the BL54L family as a whole: a scalable range of pre-certified modules and accessories that let a customer enter at whatever level of complexity fits their product today, memory and I/O with the BL54LM20A, Edge AI with the BL54LM20B, or a zero-redesign path with the USB-C Adapter, and move up that range as their product roadmap evolves, without ever losing the RF design, certification and security foundation they started with. That scalability, more than any single feature, is what we expect will drive continued adoption as more customers look to add secure, low-power Edge AI to both new and existing connected products.