M5Stack LLM-8850 Card is an M.2 M-Key 2242 AI acceleration card based on the Axera AX8850 SoC, developed to add edge devices such as the Raspberry Pi 5, RK3588 advertisement boards and x86 mini PCs. The module provides 24 TOPS (INT8) NPU throughput and has a very broad support of AI workloads including CNNs, Transformers, LLMs, and multimodal models. Its x2 size PCIe 2.0 interface allows it to connect with other components to be expanded and become a powerful local AI platform capable of real-time inference and video analytics.
Incorporated into the card is an octa-core Cortex-A55 1.7 GHz CPU, 8GB LPDDR4x memory and a hardware VPU capable of supporting 8K H.265/H.264 encoding and decoding with a capability of supporting up to 16 channels of 1080p video streams. It has a built-in active cooling system with EC-controlled fan control and a real-time power chain with a stable performance under the constant high-load working condition. The following characteristics can be used to describe the applicability of the LLM-8850 Card to surveillance, Multi-camera analysis, and advanced local AI processing.
M5Stack LLM8850 card specifications:
Axera AX8850 SoC powers the M5Stack LLM-8850 Card and has a octa-core Cortex-A55 CPU at 1.7 GHz as well as a 24 TOPS INT8 NPU to provide acceleration to AI applications. It has also a hardware security engine which supports AES, DES, 3DES and SHA-256 to provide secure boot and encryption. The module contains 8GB of 64-bit LPDDR4x memory with 4266 Mbps of speed, as well as 32Mbit QSPI NOR flash allocated to bootloader.
In the case of video workloads, the VPU of the AX8850 can encode 8K at 30fps using the H.264/H.265 video codec, and decodes 8K at 60fps. It can decode as much as 16 channels of 1080p of video simultaneously and has built in scaling and cropping options to allow it to be used to process video streams with flexibility. It has connectivity capabilities such as dual Gigabit Ethernet MAC, one USB 3.0 port, and two USB 2.0 ports. Host devices are connected to an M.2 M-Key (NGFF 2242) connector supporting 2x PCIe 2.0 lanes and enabling plug-and-play expansion to other compatible SBCs and mini PCs.
The card employs a micro turbine fan coupled with CNC-machined aluminum heatsink and the temperature, current and fan speed are controlled with an onboard EC to ensure consistent thermal performance. It works off a 3.3V supply, but fills up to 7 W when loaded and is rated to run between 0 to 60°C, approximately 70°C at full load in room temperature. The small module has a size of 42.6 x 24.0 x 9.7 mm and weighs 14.7 grams and can be used in edge AI and video-analysis in environments with small enclosures.
M5stack LLM-8850 Card Overview
Only Linux distributions are compatible with the M5Stack LLM-8850 Card, with official support for Debian 12 and Ubuntu 20.04, 22.04, and 24.04. Virtualization environments such as WSL, along with Windows and macOS systems, are not supported. This Linux-only approach aligns with the AXCL Runtime architecture, which provides C and Python APIs for device interaction, model deployment, and management.
Installation of the AXCL Runtime follows standard Linux package procedures, making setup straightforward for supported systems. M5Stack’s official documentation offers detailed installation steps, example code, and complete technical references to help developers get started with AI workloads on the LLM-8850 Card.
M5stack LLM-8850 Card Compatible Model List
M5Stack sells the LLM-8850 M.2 Axera AI module on AliExpress for $166.80 and its own online store for $139.
Niladri Chowdhury