nRF54LM20B SoC
The first Nordic SoC with an integrated Axon NPU — 128 MHz Arm Cortex-M33, 2 MB NVM, 512 KB RAM, BLE 6.0 / Matter / Thread / Zigbee, and high-speed USB.
The nRF54LM20B is the AI-accelerated variant of the large-memory nRF54LM20 family. Architecturally identical to the nRF54LM20A on the wireless and CPU side, the B variant swaps the secondary RISC-V coprocessor for the Axon NPU — a dedicated neural-network accelerator.
Both variants share the same package and the same nRF54LM20 DK (PCA10184). Pick the A if you don't need acceleration and want a general-purpose RISC-V coprocessor. Pick the B if you're running ML inference and want the speedup and energy savings.
Family at a glance
| Feature | nRF54LM20A | nRF54LM20B |
|---|---|---|
| CPU | 128 MHz Arm Cortex-M33 (TrustZone, FPU, DSP) | 128 MHz Arm Cortex-M33 (TrustZone, FPU, DSP) |
| Coprocessor | 128 MHz RISC-V | 128 MHz Axon NPU |
| NVM (Flash) | 2 MB | 2 MB |
| RAM | 512 KB | 512 KB |
| GPIO | up to 66 | up to 66 |
| USB | High-speed | High-speed |
| Radio | 2.4 GHz multiprotocol | 2.4 GHz multiprotocol |
| Wireless protocols | BLE 6.0, Matter, Thread, Zigbee, Aliro | BLE 6.0, Matter, Thread, Zigbee, Aliro |
| AI accelerator | — | ✅ Axon NPU |
Wireless capabilities
The 2.4 GHz radio is the same proven multiprotocol design used across the nRF54L Series:
- Bluetooth Low Energy 6.0 — including Channel Sounding for ranging
- Matter over Thread — full IPv6 mesh
- Thread 1.4 — IPv6 mesh networking, foundation for Matter
- Zigbee 3.0 — legacy mesh interop
- Aliro — the Apple/Google/Samsung-backed digital-key standard
This is what makes the chip uniquely suited to wake-word smart speakers, intelligent door locks, voice-controlled wearables, and anomaly-detecting predictive-maintenance sensors — the kind of products where the BLE / mesh link is just as important as the on-device inference.
Memory layout
2 MB of NVM and 512 KB of RAM is generous by Cortex-M33 standards and is specifically sized to host both your Zephyr application and a quantised neural network model. As a rough rule of thumb you can plan on:
- ~1.0–1.5 MB NVM for your Zephyr application + BLE stack + drivers
- ~256–512 KB NVM for the compiled Axon model header and weights
- ~256–384 KB RAM for Zephyr runtime + BLE stack + sensor buffers
- ~128–192 KB RAM for inference activations and tensor scratch
Real numbers depend on which model you compile — see the Axon Compiler footprint reports.
Development kit
| Item | Detail |
|---|---|
| Kit name | nRF54LM20 DK |
| Part number | PCA10184 |
| Supports | nRF54LM20A and nRF54LM20B (same board, swap the SoC) |
| Sensors on board | Microphone (for KWS demos), buttons, LEDs |
| Debugger | On-board J-Link |
| Vendor link | nordicsemi.com/Products/Development-hardware/nRF54LM20-DK |
The DK is supported upstream in Zephyr (nrf54lm20dk board) and in
the nRF Connect SDK with the Edge AI Add-on installed.
Where it sits in the Nordic line-up
nRF54L15
The mainstream nRF54L Series part — same radio, smaller memory, no NPU. Pick this if your design fits in 1.5 MB Flash and you don't need ML acceleration.
nRF54LM20
The large-memory nRF54L Series part (the "A" variant in the platforms catalogue). Same 2 MB / 512 KB layout, RISC-V coprocessor, no NPU.
nRF54H20
The high-performance multi-core nRF54H Series part. More cores, more DSP, but no NPU and a higher power envelope.
Where to next
- Understand the silicon: Axon NPU architecture
- Pick a tool: Development tools
- Build your first model: Getting started