The Maix4-HAT is a compact AI HAT for the Raspberry Pi 5 built around the MAIX-IV core board, which includes AXERA’s AX650N/C octa-core Cortex-A55 CPU, SoC, delivering up to 72 TOPS (INT4) or 18 TOPS (INT8) with 8K video encode/decode support.
It features 8 GB LPDDR4x RAM and 32 GB eMMC storage, as well as interfaces such as PCIe 2.0, HDMI 2.0a (4Kp60), USB 3.0, USB 2.0, multiple camera inputs, SPI LCD, I²C touch, speaker, and fan connectors. Designed for plug-and-play use with Raspberry Pi 5 and other boards, it accelerates Transformer-based models, making it ideal for space-constrained edge AI tasks in smart cameras, industrial inspection, and multimodal AI applications.
Sipeed Maix4-HAT Specifications:
- SoC – AXera AX650N
- CPU – Octa-core Arm Cortex-A55 @ 1.7 GHz with NEON support
- NPU – 43.2 TOPS @ INT4, 10.8 TOPS @ INT8 with support for INT4, INT8, INT16, FP16, and FP32 inputs; Equivalent to NVIDIA 40 TOPS according to Sipeed/AXERA.
- ISP – Up to 8Kp30 (8,192×4,320 @ 30 fps). maximum resolution 16,384 x 16,384
- DSP – 800 MHz dual-core DSP
- Video Decoding – H.264/H.264 video decoder 8Kp60, up to 32x channels @ 1080p30
- Video Encoding – H.264/H.265 video encoder up to 8Kp30, up to 32x channels @ 1080p30
- System Memory – 8 GB 64-bit LPDDR4x (2 GB for system + 6 GB for AI)
- Storage
- 32 GB eMMC 5.1 flash
- MicroSD card support
- Display
- Mini HDMI 2.0a up to 4K @ 60 fps
- 10-pin FPC SPI display interface with a 6-pin FPC I²C touch interface
- Camera – 0.8 mm 4-pin USB camera interfaces
- Audio
- 2-pin speaker connector
- Onboard microphone
- USB
- USB 2.0 (480 Mbps) Type-C OTG port
- USB 3.0 (5 Gbps) Type-A port
- Debugging – USB Type-C port
- Expansion – 1× PCIe 2.0 (1-lane, 16-pin FPC), Raspberry Pi 5 compatibility
- Misc
- Reset, boot 0, and boot 1 buttons
- 1.25mm 2-pin cooling fan header
- Blue and White LED indicator on the core board
- Power Supply – 5V from Pi or other boards
- Dimensions – 65 x 56 mm
The M4N-HAT supports a full AI development stack and works with Sipeed’s updated MaixPy platform with proper documentation. Developers can use pre-trained AI models from Hugging Face and convert or deploy their own models using AXERA’s Pulsar2 tool. It supports popular AI tasks like image recognition, object detection, and even running large language models. The board is fully compatible with Raspberry Pi 5, with setup and usage guides available in the AXCL documentation. More information, including hardware and software documentation, SDK downloads, AI development resources (AXCL, Pulsar2), model hubs, and sample code repositories, is available on the wiki.
The benchmark charts compare the Maix4-HAT at 18 TOPS against the RK3588 with 6 TOPS NPU, Hailo-8 (26 TOPS), and Hailo-8L (13 TOPS) across various AI models. We’ve long noted that TOPS numbers provided by the manufacturers do not always transfer into real results. In raw FPS performance, the Maix4-HAT consistently outperforms all others in most CNN and Transformer workloads, reaching up to 5,961 FPS on SqueezeNet11 and 5,073 FPS on MobileNetV2.
The bottom chart shows performance as a percentage compared to the Maix4-HAT (which is set to 100% for every model). If another chip runs at 50%, it means it’s only half as fast as the Maix4-HAT for that model. In most cases, the other accelerators (RK3588, Hailo-8, Hailo-8L) are less than half as fast as the Maix4-HAT. The only big exception is ResNet50 on Hailo-8, where the Hailo module beats the Maix4-HAT at about 131% of its speed. So, overall, the Maix4-HAT is usually much faster, especially for computer vision models, but there are a few cases where another chip can be faster for a specific model type.
There are plenty of other AI HATs for the Raspberry Pi 5, including the official Raspberry Pi AI HAT+, the Pineboards AI Bundle (Hailo 8L), and the Raspberry Pi AI Kit, all of which enable you to perform on-device AI acceleration for vision, speech, and other ML tasks.
The Sipeed Maix4-HAT can now be purchased on AliExpress for $173, although it used to be on Sipeed’s AliExpress store for $149.
Debashis Das is a technical content writer and embedded engineer with over five years of experience in the industry. With expertise in Embedded C, PCB Design, and SEO optimization, he effectively blends difficult technical topics with clear communication
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.
