Sipeed has introduced the M4N-Hat, a compact AI computing module powered by the AXERA AX650N/C chip, delivering up to 72 TOPS@INT4 (18 TOPS@INT8) of performance. The board supports 8K video encoding or decoding and is designed for edge AI applications, like smart cameras, robotics, and industrial inspection.
The module is plug-and-play compatible with Raspberry Pi boards, including the Raspberry Pi 5, and accelerates Transformer-based architectures for large quantised models like QWen 2.5, QWen 3, DeepSeek, and InternVL2.5.
It includes an 8-core Cortex-A55 CPU (1.7 GHz), dual-core DSP, 8GB LPDDR4x RAM (configurable allocation), and 32GB eMMC 5.1 storage. Expansion options feature PCIe 2.0 (1 lane), HDMI 2.0a (4K@60fps), USB 5Gbps, multiple camera interfaces, and support for peripherals like displays, touch modules, and cooling fans. With its high efficiency, strong AI acceleration, and compact design, the M4N-Hat offers developers a ready-to-use solution for multi-modal edge computing nodes.
M4N-Hat AI Module specifications:
The Sipeed M4N-Hat is powered by the AXERA AX650N/C SoC, featuring an octa-core Arm Cortex-A55 CPU clocked at 1.7GHz with FPU and NEON support, alongside a dual-core DSP at 800MHz. Its integrated NPU delivers up to 72 TOPS at INT4 (18 TOPS at INT8) and supports multiple input formats including INT4, INT8, INT16, FP16, and FP32, with TopN (N≤32) support. The module is equipped with 8GB of 64-bit LPDDR4x memory (default split: 2GB for the system and 6GB for AI workloads) and 32GB of eMMC 5.1 storage for system use. For video, it supports H.264/H.265 hardware codec with up to 8K@60fps decoding and 8K@30fps encoding, and includes an HDMI 2.0a output at up to 4K@60fps.
On the connectivity side, the board provides 11x 0.8mm 4-pin USB camera interfaces, a 16-pin FPC PCIe 2.0 interface (1-lane @ 5Gbps) fully compatible with the Raspberry Pi 5, and dual USB ports — one Type-A USB 5Gbps and one Type-C USB 480Mbps. Additional I/O includes connectors for a speaker, cooling fan, SPI display, and I2C touch interface, along with onboard reset and boot buttons. The module draws 5V power from Pi boards and measures a compact 65 x 56 mm.
For your understanding of Raspberry Pi 5, you can check out the CM5 Minima career board, microSD-Express Pi Adapter and many more.
Hardware and software documentation for the M4N-Hat can be found on Sipeed’s website, with SDKs, AI toolchain docs, and model samples also available. AXERA’s Pulsar2 compiler supports ONNX conversion and deployment, while pre-trained models are hosted on Hugging Face. Source code for samples and LLMs is provided on GitHub. For kernel, OS, or SDK customisation, developers can contact support@sipeed.com.
Currently, the Maix4-HAT is available on AliExpress at $173 USD, and for information, you can check the company’s wiki page.
Thanks to CNX-Software for the tip.
Debashis Das