M5Stack AI-8850 LLM Accelerator M.2 Kit 8GB Version (AX8850)

Description

The M5Stack LLM-8850 Kit is an 8GB AI accelerator kit for edge AI and embedded computing. It combines the LLM-8850 Card, based on the Axera AX8850 SoC, with an LLM-8850 PiHat adapter board so Raspberry Pi 5 systems can add compact high-performance AI acceleration through PCIe.

This AI accelerator kit uses an M.2 M-Key 2242 card with 24 TOPS @ INT8 compute performance, 8GB 64-bit LPDDR4x memory, an octa-core Cortex-A55 CPU, hardware 8K H.264/H.265 video encode/decode and active cooling with a micro blower fan plus CNC aluminum heatsink.

Notice: Power the whole system through the USB Type-C PD input on the adapter board. The power supply must provide at least 9V@3A (27W), and power should not be supplied from the Raspberry Pi 5 side to avoid device damage.

The device may become hot during operation. Do not touch it to avoid burns.

Because Raspberry Pi internal interrupt resources are limited, when using PCIe to a 4-lane M.2 adapter, an M.2 SSD and the LLM-8850 cannot be used at the same time.

Features

• Ultra-compact M.2 M-Key 2242 AI acceleration form factor.
• Axera AX8850 SoC with 24 TOPS @ INT8 NPU and octa-core Cortex-A55 1.7GHz CPU.
• 8GB 64-bit LPDDR4x high-bandwidth memory at 4266 Mbps.
• Hardware video engine for 8K @ 30fps H.264/H.265 encoding and decoding, with 16-channel 1080P parallel decoding.
• Secure boot and encryption through AES / DES / 3DES / SHA-256 hardware security module.
• Native AXCL support for model deployment including CNN, Transformer, CLIP, Whisper, Llama3.2, Qwen3 and InternVL3.
• Active cooling with micro blower fan, CNC aluminum heatsink and EC temperature-current-speed control.
• LLM-8850 PiHat adapter board with 2.54-40P bus and USB Type-C PD power input.
• USB PD3.0 100W protocol support with DC 24V overvoltage protection.

Specifications

Applications

• Edge AI inference deployment.
• Embedded video analytics and hardware video processing.
• Raspberry Pi AI computing expansion.
• Multimodal large-model implementation.

Product List

• 1 x LLM-8850 Card (8G)
• 1 x LLM-8850 PiHat
• 1 x 2.54-40P Female Header
• 1 x Screw and Standoff Accessory Pack

Other Content

System Compatibility

• Supported systems listed in the documentation include Ubuntu 20.04, Ubuntu 22.04, Ubuntu 24.04, Debian 12, Debian 13, Windows 10 and Windows 11.
• macOS, WSL, VMware and VBox are listed as unsupported.

Hardware and Software Notes

• The LLM-8850 Card User Guide covers AI accelerator setup and software usage for the card.
• The card documentation states that compatible hosts include Raspberry Pi 5, RK3588 SBCs and x86 PCs with suitable PCIe M.2 M-Key support.
• For the card interface, PCIe 2.0 x2 is supported and backward compatibility with x1 is noted, but NVMe-protocol M.2 SSD interfaces are not supported.

Resources

• LLM-8850 Kit Product Documentation
• LLM-8850 Card Product Documentation
• LLM-8850 Card User Guide
• LLM-8850 Kit Schematic PDF
• LLM-8850 Card Model Size PDF
• LLM-8850 PiHat Adapter Board Model Size PDF
• Axera AX8850 Product Brief

FAQ

What is the M5Stack LLM-8850 Kit used for?

It adds edge AI acceleration to host systems such as Raspberry Pi 5, helping with multimodal model inference, video analytics and embedded AI workloads.

How much AI performance does the AI accelerator kit provide?

The Axera AX8850 NPU provides up to 24 TOPS @ INT8 compute performance.

How much memory does this version include?

This 8GB version uses 64-bit LPDDR4x memory at 4266 Mbps in a 4+4 configuration.

How should the system be powered?

Use the USB Type-C PD input on the adapter board with a power supply of at least 9V@3A (27W). Do not power it from the Raspberry Pi 5 side.

Can it be used together with an M.2 SSD on Raspberry Pi?

When using PCIe to a 4-lane M.2 adapter, the documentation says an M.2 SSD and the LLM-8850 cannot be used simultaneously because of limited Raspberry Pi interrupt resources.

Is cooling built in?

Yes. The LLM-8850 Card includes active cooling with a micro blower fan and CNC aluminum heatsink, controlled by onboard EC thermal logic.