LLM‑8850 Card is an M.2 M-KEY 2242 AI accelerator card designed for edge devices. It combines a compact 42 mm form factor with the Axera AX8850 SoC delivering 24 TOPS @ INT8 compute performance, enabling “plug-and-boost” multimodal large model and video analytics acceleration for hosts such as Raspberry Pi 5, RK3588 SBCs, and x86 PCs. The card integrates an active cooling system consisting of a micro centrifugal fan and a CNC aluminum heatsink. Fan speed is intelligently regulated by the onboard EC based on temperature–current curves, ensuring stable low-temperature operation even under sustained full load and preventing thermal throttling in enclosed chassis environments.
The onboard DCDC + PMIC power architecture is managed in real time by the EC to achieve “power on demand, cooling on demand,” significantly improving overall system stability. It supports AXCL Runtime with C / Python APIs for one-click deployment of mainstream CNN, Transformer, LLM, and multimodal models such as YOLO‑v8/11, CLIP, Whisper, Llama3.2, InternVL3, and Qwen3. Leveraging the AX8850 VPU hardware pipeline, it also provides H.264/H.265 8K video encoding/decoding, simultaneous encode–decode transcoding acceleration, and scaling / cropping. AI inference and video stream processing can be handled concurrently, with host-side support for directly invoking the hardware video codec via ffmpeg.
Native AXCL support: one-click execution of CNN, Transformer, CLIP, Whisper, Llama3.2, Qwen3, InternVL3, and more, with H.264/H.265 simultaneous encode–decode transcoding
Includes
1 x LLM‑8850 Card
Applications
Industrial / commercial SBC compute upgrades: local object detection and defect inspection on Raspberry Pi 5, RK3588, TI AM62x, and similar boards
Embodied intelligence robots: plug-and-play local “perception–decision–control” pipelines for AMR / AGV / service robots
AIPC & edge intelligent terminals: offline Copilot, customer service Q&A, meeting subtitles, and real-time translation inside mini PCs
NVR / NAS intelligent upgrades: add multi-channel AI license plate recognition, event summarization, and hardware transcoding to legacy storage devices
Intelligent interactive devices: low-latency local LLM + TTS for voice assistants, smart doorbells, and digital signage
AI vision gateways: real-time inference for traffic intersections and campus access control, including people flow statistics and hazardous behavior alerts
Micro centrifugal fan + CNC aluminum integrated heatsink, EC smart thermal control
Operating Temperature
0 ~ 60 °C
Full-load Temperature (Ambient)
70 °C
Power Supply
7W @ 3.3V
Product Size
42.6 x 24.0 x 9.7mm
Product Weight
14.7g
Package Size
66.0 x 44.0 x 13.5mm
Gross Weight
19.8g
Learn
Reminder
The device generates heat during operation. Do not touch it to avoid burns.
Power Requirement
When connecting to a Raspberry Pi or other PC, please use a switching power supply adapter capable of DC 5V@3A (non-PD protocol). If a PD power adapter is used, protocol negotiation issues may prevent full power output, potentially causing abnormal device operation.
Device Requirements
Due to limited internal interrupt resources on the Raspberry Pi, when using a PCIe to quad-lane M.2 adapter, M.2 SSDs cannot be shared with the LLM‑8850 Card. Currently, the Waveshare PCIe to dual-lane M.2 adapter is known to be unsupported.
Form factor requirement: PCIe M.2 M-Key
Interface requirement: Supports PCIe 2.0 ×2 lanes, backward compatible with x1 (e.g., Raspberry Pi PCIe 2.0 x1). Note: NVMe-protocol M.2 SSD interfaces are not supported.
Hardware requirement: When used with Raspberry Pi, the official Raspberry Pi controller and adapter board or an M5Stack adapter board is recommended. Third-party hardware may have compatibility issues.
System requirement: Refer to the system compatibility table below. Other systems may not install drivers correctly.
Hardware Compatibility
The following boards have been tested by official and third-party users and work properly with the LLM‑8850 Card: