













Description
LLM‑8850 Card is an M.2 M-KEY 2242 AI acceleration card designed for edge devices, combining a compact 42 mm form factor with the Axera AX8850 SoC delivering 24 TOPS @ INT8 performance. It enables host devices such as Raspberry Pi 5, RK3588, and x86 PCs to easily expand multimodal large-model and video analysis capabilities with a simple plug‑and‑play design. The card is equipped with an active cooling system featuring a micro turbine fan and aluminum alloy CNC fins, with fan speed intelligently controlled by the onboard EC based on temperature–current curves. Even under long‑term full load, it can maintain stable low temperatures to prevent thermal degradation inside enclosed cases.
The onboard DCDC + PMIC power chain is managed in real time by the EC to achieve “power on demand, cooling on demand,” greatly improving overall system stability. Supports AXCL Runtime, with C / Python APIs enabling one-click deployment of mainstream CNN, Transformer, LLM, and multimodal models such as YOLO‑v8/v9, CLIP, Whisper, Llama 3.2, Gemma 2, Qwen 2.5; and utilizes the AX8850’s VPU hardware pipeline for H.264/H.265 8K encoding/decoding with simultaneous encode-decode transcoding acceleration and scaling/cropping, bridging AI and video stream processing. Host devices can directly invoke the hardware video codec via ffmpeg.
Features
- Ultra-compact form factor: NGFF M.2 M-KEY 2242 size, supports PCIe 2.0 ×2 lanes plug-and-play
- High-performance NPU: 24 TOPS @ INT8, octa-core Cortex‑A55 1.7 GHz CPU
- Intelligent cooling / power supply: onboard turbo fan + CNC aluminum alloy integrated heatsink, EC-monitored temperature-current-speed closed loop
- High-bandwidth memory: 64‑bit LPDDR4x, 4266 Mbps speed, 8GB capacity
- Rich I/O: 1 × USB 3.0, 2 × USB 2.0, 2 × Gigabit Ethernet MAC
- Hardware video engine: 8 K @ 30 fps H.264/H.265 encoding, 8 K @ 60 fps decoding, supports 32-channel 1080 p parallel decoding
- Secure boot & encryption: AES / DES / 3DES / SHA‑256 hardware security module
- Native AXCL: one-click operation for CNN, Transformer, CLIP, Whisper, Llama3.2, Qwen 2.5, InternVL 2 full-stack models, supports H.264/H.265 simultaneous encode-decode transcoding
Includes
- 1 x LLM‑8850 Card
Applications
- Industrial / commercial SBC compute upgrade: local execution of object detection and defect recognition on Raspberry Pi 5, RK3588, TI AM62x boards
- Embodied intelligent robots: AMR / AGV / service robots with plug-and-play local “perception-decision-control” chains
- AIPC & edge intelligent terminals: provide offline Copilot, customer service Q&A, meeting subtitles & instant translation within mini PCs
- Intelligent NVR / NAS upgrades: adding multi-channel AI license plate capture, event summary search + hardware transcoding to existing storage boxes
- Smart interactive devices: enable low-latency local LLM + TTS dialog for voice assistants, smart doorbells, advertising displays
- AI vision gateways: real-time passenger flow statistics and dangerous behavior warnings at traffic intersections and park gates
Specifications
Specification | Parameter |
---|---|
SoC | Axera AX8850 |
CPU | Octa-core Cortex‑A55 1.7 GHz |
NPU | 24 TOPS @ INT8 |
Video Encoder | 8 K @ 30 fps H.264/H.265 encoding, supports scaling / cropping |
Video Decoder | 8 K @ 60 fps H.264/H.265 decoding, supports 32 channels 1080 p parallel decoding, supports scaling / cropping |
Memory | 64‑bit LPDDR4x, 4266 Mbps, 8GB capacity |
Storage | 32Mbits QSPI NOR Flash (used for Bootloader only) |
Form Factor | M.2 M-KEY 2242, PCIe 2.0 ×2 |
Cooling | Micro turbo fan + integrated aluminum alloy CNC heatsink, EC intelligent temperature control |
Operating Temperature | 0 ~ 60 °C |
Full Load Temperature at Room Temp | 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 |