NPU for the Raspberry Pi 5: Edge AI accelerator delivers up to 40 Tops
Raspberry Pi Plc. releases the PCI Express module AI HAT+ 2 for the Raspi 5. It combines the AI chip Hailo-10H with 8 GB LPDDR4X RAM for object recognition.
Raspberry Pi AI HAT+2 with Hailo-10H on a Raspberry Pi 5
(Image: Raspberry Pi)
The Raspberry Pi 5 has a proprietary PCI Express connector for attaching expansion boards. The AI accelerator Raspberry Pi AI HAT+ 2 also docks onto this. It is equipped with the AI coprocessor Hailo-10H, which in turn connects 8 gigabytes of LPDDR4X RAM.
The list price of the AI HAT+ 2 is initially $130 due to high RAM prices. Experience shows that the price in Germany is likely to be over 140 Euros.
Videos by heise
Predecessor AI HAT+
In June 2024, Raspberry Pi Plc. introduced the predecessor AI HAT+, which is available in two versions: either with Hailo-8 and 26 trillion operations per second (Tops) or with Hailo-8L and 13 Tops.
The Hailo-10H on the AI HAT+ 2 processes Int8 data values with 20 Tops, but can also handle Int4, achieving 40 Tops. For comparison: Microsoft requires a Neural Processing Unit (NPU) with 40 Int8 Tops for processors for Windows 11 notebooks that are allowed to bear the "Copilot+" logo. The NPU in the Apple M4 delivers 38 Int8 Tops.
According to Hailo, the power consumption for both Hailo-8 and Hailo-10H is 2.5 watts each. However, the Hailo-10H also requires the separate RAM. Raspberry Pi Plc. advises paying attention to adequate cooling.
Model Zoo
(Image:Â Raspberry Pi/Hailo)
According to Raspberry Pi, a Raspi 5 with AI HAT+ 2 should be able to run certain AI language models (LLMs) locally. However, the system is more likely to be used for typical edge AI applications, such as the analysis of video images.
The drivers for the AI HAT+ are integrated into Raspberry Pi OS. A Raspberry Pi camera connected via MIPI CSI can also be integrated.
Additionally, examples of AI models for the Hailo chips are available for trying out as well as templates for your own developments.
(ciw)
