$70 Raspberry Pi AI Kit combines official M.2 HAT+ with Hailo-8L AI accelerator

Raspberry Pi Limited has just launched the “Raspberry Pi AI Kit” comprised of the official M.2 Key M HAT+ and a 13 TOPS Hailo-8L M.2 AI accelerator module and selling for $70 through distributors.

We had seen Raspberry Pi showcase an AI camera at Embedded World 2024, so when I received an email from a representative about a “Raspberry Pi AI Kit” I thought it would be the announcement about the camera. Instead, it’s a kit comprised of existing parts with the most interesting aspects being the price and availability (hopefully) since Hailo-8/8L accelerators are mostly found in more expensive embedded/industrial solutions, and easier documentation to get started.

Raspberry Pi AI Kit Raspberry Pi 5

Raspberry Pi AI Kit highlights:

  • Supported SBC – Raspberry Pi 5
  • M.2 HAT+ with PCIe Gen2 x1 interfaces, M.2 Key M support,
  • Hailo-8L AI accelerator with
    • Up to 13 TOPS of performance
    • M.2 2242 form factor
    • Typical power consumption – 1.5W
  • Thermal pad pre-fitted between the module and the M.2 HAT+
  • Mounting hardware kit
  • 16mm stacking GPIO header
  • PCIe FPC cable

Raspberry Pi AI Kit object detection

There’s nothing really special about the hardware, but Raspberry Pi and Hailo collaborated on the software and you’ll find a range of AI-accelerated computer vision samples for the Raspberry Pi 5, including object detection, pose estimation, and instance segmentation on GitHub, along with documentation on the Raspberry Pi website.

The main advantage of using an AI accelerator such as the Hailo-8L over the CPU or GPU on a Raspberry Pi 5 is the lower power consumption and much faster AI processing as can be seen on the pose estimation demo shown in the video embedded below.

More sample programs leveraging the new Raspberry Pi AI kit are coming soon notably a CLIP (Contrastive Language-Image Pretraining) program that predicts the most relevant text prompt on real-time video frames using the Hailo-8L AI processor. Raspberry Pi also worked on Raspberry Pi rpicam-apps making use of the Pi cameras, and work is being done on picamera2 to support Hailo-8L through a Python API.

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress

ROCK 5 ITX Rockchip RK3588 mini-ITX motherboard
Subscribe
Notify of
guest
The comment form collects your name, email and content to allow us keep track of the comments placed on the website. Please read and accept our website Terms and Privacy Policy to post a comment.
7 Comments
oldest
newest
Hedda
Hedda
6 months ago

Support for OneAPI?

persondb
persondb
6 months ago

Wouldn’t the slow connection be a bottleneck? The Hailo-8L doesn’t have it’s own external memory pool, only what is on the chip and that should mean a lot of weights and data would be forced to be in the Host memory, which can only be streamed through the slow PCIe Gen 2 x1.

Hedda
Hedda
6 months ago

Yeah the memory size will be the bottleneck, as with more memory you could at least use larger models even if it would be slow, but without more memory you can not even load larger models

persondb
persondb
6 months ago

I meant it in the sense that a lot of those accelerators can load part of the model in their on-chip memory and execute that. This is likely what is going to happen with those small accelerators without external memory. Google Coral TPU does that.

Then that would become a bottleneck due to the low connection speed from the RPi to the acelerator for the streaming of the data.

foxnux
foxnux
6 months ago

Right. I guess there should be at least two PCIe 3.0 lanes to get over 10 TOPS in the real world complicated YOLO real-time object detection.

Kowskirz
6 months ago

Jeff Geerling built a 50+ TOPS Raspberry Pi “ML” model that he called “CoPilot” (Geddit, Pi… lot). Kind of cool.

Boardcon Rockchip RK3588S SBC with 8K, WiFI 6, 4G LTE, NVME SSD, HDMI 2.1...