Specifications
Brand
Sipeed
Model
Maix4-HAT (M4C-Hat)
Accelerator
Axera AX650N
CPU
8× Arm Cortex-A55 @ 1.7 GHz with NEON
NPU
72 TOPS INT4 / 18 TOPS INT8 — INT4, INT8, INT16, FP16, FP32 inputs
DSP
Dual-core @ 800 MHz
Memory
8 GB 64-bit LPDDR4x (default 2 GB system + 6 GB AI)
Storage
32 GB eMMC 5.1 system storage
Video Codec
H.264 / H.265, up to 8Kp60 decode and 8Kp30 encode
Video Output
HDMI 2.0a up to 4Kp60
Host Interface
PCIe, HAT form factor for Raspberry Pi 5
Framework Support
ONNX and standard edge inference pipelines
Workload Note
SmolVLM-256M image encoder in ~105 ms; Stable Diffusion 1.5 U-Net at 0.43 s/iteration (vendor figures)
Overview
The Sipeed Maix4-HAT adds a dedicated AI accelerator to a Raspberry Pi 5 over PCIe. Its AX650N silicon carries a 72 TOPS INT4 / 18 TOPS INT8 NPU alongside eight Cortex-A55 cores at 1.7 GHz and a dual-core DSP, so the board runs inference and video handling itself rather than leaning on the Pi's CPU.
Memory is split deliberately: 8 GB of 64-bit LPDDR4x is allocated by default as 2 GB for the system and 6 GB dedicated to AI, which keeps model and activation storage off the host. Onboard 32 GB eMMC 5.1 holds the system image independently of the Pi's own storage.
That budget is what allows generative and vision-language workloads, not just detection. Vendor benchmarks put a SmolVLM-256M image encoder at roughly 105 ms and Stable Diffusion 1.5's U-Net at about 0.43 s per iteration — work that is impractical on a stock Pi 5 CPU.
Video support covers H.264/H.265 decode to 8Kp60 and encode to 8Kp30, with HDMI 2.0a output up to 4Kp60 directly from the module. For analytics and transcoding jobs the HAT can ingest and re-encode camera streams while the NPU runs detection on the same device.
Key Benefits
72 TOPS INT4 / 18 TOPS INT8 lifts a Raspberry Pi 5 to workstation-class edge inference · 6 GB of dedicated AI memory keeps model footprint off the host · onboard 32 GB eMMC and HDMI output allow the HAT to run standalone · PCIe HAT form factor preserves the Pi's GPIO and camera interfaces.
Applications
Generative and vision-language inference at the edge, multi-camera video analytics, 8K video transcoding and NVR duty, robotics perception on Raspberry Pi platforms, and edge AI prototyping where a Pi 5 is the base platform.
Request a Quote — SIPEED MAIX4-HAT — 72 TOPS PCIE AI HAT FOR RASPBERRY PI 5
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →