Sixfab Introduces AI HAT+ for Raspberry Pi 5 Featuring DEEPX DX-M1 AI Accelerator

Sixfab has officially launched its new AI HAT+ designed specifically for the Raspberry Pi 5. This new PCIe HAT+ leverages the DEEPX DX-M1 AI accelerator, a chip previously seen in devices like the DEEPX DX-AIPlayer, Mini DX-M1 SoM, and ALPON X5.

Unlike other hardware implementations (such as the M.2 module variant found in the ALPON X5), Sixfab's AI HAT+ features the accelerator soldered directly onto the board. It connects seamlessly to the Pi 5 using a PCIe FFC (flexible flat cable) and draws its operating power straight from the 40-pin GPIO header. The board is designed to execute localized vision-based AI tasks on the Raspberry Pi 5, including real-time object detection and image segmentation. It comes in two performance options: a 13 TOPS variant and a 25 TOPS variant.

Technical Specifications

  • Compatible SBC: Raspberry Pi 5
  • AI Accelerator Options:
  • DEEPX DX-M1M: Offers up to 25 TOPS (INT8) processing power, equipped with 1 GB LPDDR4X NPU memory.
  • DEEPX DX-M1ML: Offers up to 13 TOPS (INT8) processing power, equipped with 512 MB LPDDR4X NPU memory.
  • Host Interface: PCIe Gen 3 x1 via a 16-pin FFC cable.
  • Cooling: Passive cooling by default; includes an integrated 2-pin JST fan connector for optional active cooling.
  • Power Setup:
  • Input: 5V / 3A supplied through the Raspberry Pi 5’s 40-pin GPIO header (no external power connectors needed). Note: A 27W Power Supply Unit (PSU) is required for operation; a standard 15W PSU will not suffice.
  • NPU Peak Consumption: 2.5 to 3 W during full inference workloads.
  • NPU Idle Consumption: 0.5 to 1 W.
  • Combined System Power (Pi 5 + HAT+): 13 to 15 W.
  • Form Factor: 65 x 56.5 mm (fully compliant with Raspberry Pi HAT+ standards), with a height of 6.56 mm.
  • Operating Temperature: 0 to 70°C (Commercial grade).
  • Certifications: CE, FCC, UKCA, RoHS, REACH (currently in progress).

Software Stack and Compatibility

The AI HAT+ is fully compatible with Raspberry Pi OS (Trixie). Thanks to its integrated HAT+ EEPROM, the board supports automatic configuration. Setting it up is straightforward: users just need to install the dxrt-runtime package directly from Sixfab's APT repository, which deploys both the necessary drivers and the runtime environment.

Developers can jump straight into projects using pre-compiled AI models available from the Sixfab Model Zoo—featuring popular options like YOLOv8, MobileNet, and ResNet. Alternatively, developers can run custom models by exporting them to the ONNX format and compiling them into the native DXNN format via the DX-COM utility tool. The runtime natively supports development in both Python and C++.

Comparative Capabilities

While the Sixfab AI HAT+ delivers specialized performance for vision-based AI tasks similar to the Hailo-8-powered Raspberry Pi AI HAT+, it is not designed to run Large Language Models (LLMs) or Generative AI workloads. This limitation stems from its architecture, which lacks transformer decoder support, as well as its limited on-board memory.

In comparison, the higher-end Raspberry Pi AI HAT+ 2 (utilizing the Hailo-10H) hits up to 40 TOPS and packs 8GB of dedicated memory specifically to handle LLM and VLM applications. Sixfab has noted that LLMs are on DEEPX’s future silicon roadmap, and they intend to support generative AI workloads as hardware updates allow, though no official dates have been announced.

Pricing and Future Expansion

The Sixfab AI HAT+ with the DEEPX DX-M1 accelerator is available for purchase now on the official Sixfab store. The 13 TOPS (DX-M1ML) variant costs $63, while the 25 TOPS (DX-M1M) variant is priced at $90.

Additionally, Sixfab revealed they are developing an upcoming Edge AI Expansion Board for the Raspberry Pi 5. This upcoming multi-functional board is expected to bundle AI acceleration alongside NVMe SSD storage and LTE/5G cellular connectivity in a single form factor, though further specifics have not yet been disclosed.

Back to blog