Sixfab AI HAT+ Delivers Up to 25 TOPS Edge AI for Raspberry Pi 5
Powered by DEEPX NPUs, this new expansion board provides local vision AI processing over PCIe, eliminating cloud dependency while keeping power consumption low.
Sixfab has introduced the AI HAT+ for Raspberry Pi 5, an expansion board based on DEEPX DX-M1 series AI accelerators for computer vision and local inference. It adds on-device AI processing to the Raspberry Pi 5, removing the need for external GPUs or cloud-based processing.
Unlike the M.2-based AI modules found in other Edge AI platforms like the ALPON X5, the AI HAT+ features the neural processing unit (NPU) soldered directly onto the PCB. It interfaces with the Pi via its PCIe Gen 3 x1 port using a standard 16-pin FFC cable, which bypasses any USB bandwidth bottlenecks. Power is drawn straight from the Pi’s 40-pin GPIO header, though a 27W power supply is required to handle peak loads. The board comes in two versions, which include a 13 TOPS-based model powered by DEEPX DX-M1ML and a 25 TOPS model based on the DX-M1M. The HAT can be used for tasks like object detection, segmentation, and image classification.
Sixfab AI HAT+ top view
Sixfab AI HAT+ Specifications:
- Host SBC: Raspberry Pi 5
- AI Acceleration Options:
- DEEPX DX-M1M: Up to 25 TOPS (INT8) with 2 GB LPDDR4X memory
- DEEPX DX-M1ML: Up to 13 TOPS (INT8) with 1 GB LPDDR4X memory
- Interface: PCIe Gen 3 x1 via 16-pin FFC cable
- Power:
- Supply: 5V / 3A via the Raspberry Pi 5 40-pin GPIO header (27W PSU recommended; standard 15W is insufficient)
- Consumption: ~3W NPU peak draw (~13–15W combined Pi 5 + HAT+ under full load)
- Cooling: Passive by default; includes an onboard 2-pin JST connector for an optional fan
- Form Factor: Raspberry Pi HAT+ compliant (65 x 56.5 x 6.56 mm)
- Operating Temperature: 0°C to 70°C (Commercial grade)
Sixfab AI HAT+ installed on Raspeberry 5
On the software side, the board follows the Raspberry Pi HAT+ standard and uses EEPROM for automatic setup. It works on Raspberry Pi OS with a runtime package (dxrt-runtime). Pre-built models like YOLOv8, MobileNet, and ResNet can be used, and custom models can be converted using ONNX and compiled with the DEEPX toolchain. It supports both Python and C++ for development.
The AI HAT+ is designed for computer vision tasks and is similar in performance to the Hailo-8 AI HAT+, but it is not suitable for running generative AI or large language models due to hardware limitations. Support for such workloads may be added in future versions of the DEEPX chips.
The board is available now from the Sixfab store, with the 13 TOPS version priced at $63 and the 25 TOPS version priced at $90. More information about the board can be foudn on there docoumentations page.
Images used courtesy of Sixfab
