2025-07-10

Introduction

The M4N-Hat is a highly integrated AI computing module developed by Sipeed, featuring a compact design powered by AXERA's AX650N/C chip. As an embedded expansion module, it delivers 72 TOPS@INT4 (18 TOPS@INT8) computing power while retaining 8K video encoding/decoding capabilities, making it ideal for space-constrained edge computing applications.

This module supports plug-and-play compatibility with mainstream Raspberry Pi development boards and accelerates Transformer-based architectures. With its onboard 0.8mm 4-pin USB and Type-A USB SS 5Gbps interfaces, it enables quick expansion for peripherals like cameras and storage, making it perfect for smart cameras, industrial inspection, and other lightweight AI applications. The PCIE 2.0 expansion interface is fully compatible with Raspberry Pi 5, supporting multi-modal edge computing nodes and the deployment of quantized large models such as QWen 2.5, QWen 3, DeepSeek, and InternVL2.5.

Showcase

Interface Annotation Diagram

Interface Annotation Diagram

Key Specifications

Component Description
CPU 8x A55@1.7Ghz, integrated FPU, supports NEON acceleration
NPU 72 TOPS@INT4 / 18 TOPS@INT8, supports INT4/INT8/INT16/FP16/FP32 inputs, TopN (N<=32)
CODEC H.264/H.265 encoding/decoding, up to 8K@60fps decoding & 8K@30fps encoding
DSP Dual-core 800MHz
RAM 8GB 64-bit LPDDR4x (adjustable allocation: default 2GB system + 6GB AI CMM)
ROM 32GB eMMC 5.1 (system storage)
Video Output 1x HDMI 2.0a (max 4K@60fps)
Video Input 11x 0.8mm 4-pin USB camera interface
PCIE 1x 16-pin FPC (1-lane PCIE2.0 @5Gbps), Raspberry Pi 5 compatible
USB 1x Type-A USB SS 5Gbps + 1x Type-C USB HS 480Mbps
Others 1x 1.25mm 2-pin speaker, 1x 1.25mm 2-pin fan, 1x 10-pin FPC SPI display, 1x 6-pin FPC I2C touch

Performance Benchmarks


Models RK3588@6T Maix4@18T Hailo8 26T Hailo8 13T
Inceptionv1 43 2494 928 519
MobileNetv2 960 5073 2433 1738
SqueezeNet11 694 5961 - -
ResNet18 543 2254 - -
ResNet50 294 1045 1368 503
SwinT 21 401 - -
ViT-B/16 18 207 107 40
YOLOv5s 48 384 364 168
YOLOv5n 78 743 - -
YOLOv6s 80 321 - -
YOLOv6n 212 743 - -
YOLOv8s 39 279 - -
YOLOv8n 73 710 - -
YOLOxs 34 304 - -
YOLO11s 30 313 - -
Models Item Maix4@18T RK3588@6T
SmolVLM-256M Image Encoder 512*512 105ms 842ms
TTFT 57ms 87ms
Decode 80 tokens/s 77 tokens/s
StableDiffusion 1.5(512*512) U-Net 0.43 s/it 5.65 s/it
VAE Decoder 0.91 s 11.13 s
Qwen2.5-VL-3B Image Encoder 448*448 780 ms
TTFT 320 tokens 2857 ms
Decode 6.2 tokens/s
Image Encoder 392*392 2930 ms
TTFT 196 tokens 1262 ms
Decode 8.6 tokens/s

Resources

Hardware Documentation

Datasheet: https://dl.sipeed.com/shareURL/MaixIV/M4N-Dock

Software Documentation

Docs: https://dl.sipeed.com/shareURL/MaixIV/M4N-Dock
SDK: https://www.ebaina.com/down/240000038900

AI Development

LMM & AXCL: https://axcl-docs.readthedocs.io
Raspberry Pi 5 AXCL Guide: https://axcl-pi5-examples-cn.readthedocs.io

Model Hub: https://huggingface.co/AXERA-TECH

AI Toolchain (ONNX Conversion/Deployment)

  • Pulsar2 (AXERA's ALL-IN-ONE Neural Network Compiler):

Docs: https://pulsar2-docs.readthedocs.io/en/latest/pulsar2/introduction.html
Download: https://pan.baidu.com/s/1FazlPdW79wQWVY-Qn--qVQ?pwd=sbru

op_support_list: https://pulsar2-docs.readthedocs.io/en/latest/appendix/op_support_list_ax650.html

Samples source: https://github.com/AXERA-TECH/ax-samples
LLM source: https://github.com/AXERA-TECH/ax-llm

Technical Support

For custom development (kernel/OS customization, application-layer SDKs), contact: support@sipeed.