Raspberry Pi AI Kit available now at $70
If you’ve ever wanted to experiment with the world of neural networks, artificial intelligence and machine learning on your Raspberry Pi 5, we have the perfect product for you: the Raspberry Pi AI Kit. Developed in collaboration with Hailo, the AI Kit offers an accessible way to integrate local, high-performance, power-efficient inferencing into a wide variety of applications. It’s available today from our network of Raspberry Pi Approved Resellers, priced at just $70.
The Raspberry Pi AI Kit comprises our M.2 HAT+ preassembled with a Hailo-8L AI accelerator module. Installed on a Raspberry Pi 5, the AI Kit allows you to rapidly build complex AI vision applications, running in real time, with low latency and low power requirements. State-of-the-art neural networks for object detection, semantic and instance segmentation, pose estimation, and facial landmarking (to name just a few) run entirely on the Hailo-8L co-processor, leaving the Raspberry Pi 5 CPU free to perform other tasks.
Key features of the Raspberry Pi AI Kit include:
- 13 tera-operations per second (TOPS) of inferencing performance
- Single-lane PCIe 3.0 connection running at 8Gbps
- Full integration with the Raspberry Pi image software subsystem
- Compatibility with first-party or third-party cameras
- Efficient scheduling of the accelerator hardware: run multiple neural networks on a single camera, or single/multiple neural networks with two cameras concurrently
Hailo has created an extensive model zoo, where users can find a wide variety of pre-trained neural network models ready to deploy and optimised to run on the AI Kit.
The software
A significant hurdle in creating real-world AI-based vision applications is the software complexity of integrating the camera subsystem with the AI framework. We have worked hard to simplify this as much as possible. Our rpicam-apps suite of camera applications now has a post-processing template for integrating neural network inferencing running real-time in the camera pipeline. By using the pre-installed Hailo Tappas post-processing libraries, we are able to create advanced AI-based applications in only a few hundred lines of C++ code. Similar levels of integration into our Picamera2 framework will follow soon.
The software installation steps are very simple. Install a few packages through apt, reboot, and you are ready to try out some of our AI demos in a matter of minutes. The full set of instructions can be found in our getting started guide. Here’s a preview of some of our demos that you can run through rpicam-apps:
With the Raspberry Pi AI Kit, you are not limited to using the Hailo-8L co-processor only in rpicam-apps or Picamera2. We also package an API integrated in the GStreamer framework and native Python or C/C++ applications. This also includes non-camera use cases, such as running inference on pre-recorded video files.
Further resources
Our documentation for the AI Kit is a great place to start.
For full technical specifications for the Hailo-8L AI accelerator module, visit Hailo’s product web page.
Hailo has created a set of advanced AI applications running on a Raspberry Pi 5. They also have a community forum for discussing topics specific to the Hailo-8L AI accelerator hardware and software tooling.
49 comments
IanS
Is the AI module likely to work on other M.2 baseboards? In particular, could it be used as well as an SSD on a dual board such as https://shop.pimoroni.com/products/nvme-base-duo-for-raspberry-pi-5 ?
Adam
This is the first thing that occurred to me, that they often forget that the Rpi 5 only has 1 PCIE port, which NVME drives also work with – which means that the operation with it – is also the ‘expectation of our days’ and when they bring out a new unit that also works on the PCIE channel, then what about using NVME drive? ….
barleyguy
As you might expect, Jeff Geerling released a video today using multiple AI modules on PCI-E splitters.
Jeff Geerling
And I’ve already found a bug in that configuration, and filed an issue in the Pi firmware repository :)
https://github.com/raspberrypi/firmware/issues/1895
Hopefully we can get that fixed soon, I would love to use this with an NVMe SSD on a dual-NVMe board for Frigate!
Rajesh K
Have you tried with a SSD CARD and Halo AI board together ?
Jacek Ciuba
Can’t wait to get it to the test!
Raspberry Pi Staff Ashley Whittaker
What’re you thinking of doing with it? Or hoping inspiration will strike upon delivery?
Jeff Geerling
Right now it looks like Hailo’s community forum is stuck behind a registration wall. Is there any chance Hailo could open it up read-only to the public, so not only could people see the valuable information locked inside without registering, but search engines could also index that info so we can find it when debugging?
Otherwise I feel like most users will just ignore that support community and stick to the Pi forums or elsewhere.
Raspberry Pi Staff Ashley Whittaker
There are plans to open their community for read only access without registration. Stand by.
Jeff Geerling
Hooray! I know something like that can take time (and not just on the technical level), so thanks for helping them along that path :)
joaquin
I have my Pi5 running from an NVMe that uses the PCIe connector. I think this is a common setup for many people. How can be the AI board connected in this case?
Joseph Alway
I’ve seen a couple of dual m.2 base boards, but who knows if an NvME drive and the AI module would play nice together. Due to certain limitations the dual m.2 base board that I was seriously looking at couldn’t directly boot from the attached drives. Still, it might be possible to have a NvME drive and AI module in use at the same time using a dual m.2 board.
Bret
Looks good! Though I have to ask the question, does the M.2 HAT+ bundled with these AI kits come with the updated screws/spacers? Early reviews seem to have the plastic ones but I’m not sure if those were just early samples where there wasn’t time to get the updated ones out?
Raspberry Pi Staff Ashley Whittaker
Yes! Have double checked and all HATs in the AI Kits have the replacement screws. Every M.2 HAT+ going out after that first batch does as well.
Bret
Super late reply but that’s great, thanks! Would you be able to confirm if the replacement screws are black nylon/plastic still? Or are they metal? Thanks again
Anders
Is the M.2 module available by itself? If you already have M.2 HAT+, it’s not like you can use it at the same time. Or is it capable of stacking with longer GPIO adapter?
Or it is reason to buy another Pi5 and PSU etc. Very vertical selling.
Yield Estimator
Cool! This will be a really good tool to use for our yield estimation hardware & software platform! Thanks Raspberry Pi!
Julian
As a beginner I’m looking for something like this to be able to run a Whisper model on my RPI 5. Would this product help speed this prosess up?
Gordon77
Looks interesting.
I assume first batch already sold out, as UK suppliers appear to have no stock 😒
Raspberry Pi Staff Helen Lynn
Shipments are on their way to Approved Resellers everywhere as I type – you shouldn’t have long to wait.
Gordon77
Thanks. One supplier now saying batch 3 in August.
Obviously very popular 😀
Anders
I put a notify me on with the Haverhill lot, and they got their stock in already today. I’ve ordered mine, there were 16 left when I got my order in.
Raspberry Pi Staff Ashley Whittaker
Come back and tell us what you build with it, Anders!
Pete Chown
Can this be used as a graphics accelerator as well? People use graphics cards for AI, so why not use an AI accelerator for graphics? 🙂
This is a massive oversimplification so don’t take it too seriously—but a PS5 is supposed to provide about 10TFLOPS. This card is supposed to provide 13TOPS. If a large proportion of those can be single-precision float, its power as a graphics accelerator could be close to a modern console.
Warren
any ideas how much memory is on this thing? all websites I found for it seem to omit this info
Rahul Kumar
What about Hailo-10H M.2 module ? Is it also compatible with RPi 5 ?
https://hailo.ai/products/generative-ai-accelerators/hailo-10h-m-2-generative-ai-acceleration-module/#hailo10m2-features
crumble
+1
I would like to have the version with on board RAM as well ;)
The 8GB shall be enough to run whisper or a small llama… if someone can convert the models in their format.
John
Hi @Rahul
Did you manage to find out any more info on the Hailo-10H compatibility for pi 5?
Ric
Well done!
Can you create and train ANNs or only run them
Hedda
Official support for support for OneAPI planned?
https://www.oneapi.io
Kristof
Is the Hailo included in the kit?
Raspberry Pi Staff Helen Lynn
Yes, the Hailo module comes pre-assembled with the Raspberry Pi M.2 HAT+.
Andre Costa
I don’t see the module for sale separately. Looks like Hailo doesn’t do retail. It would be great to be able to buy the module separately to use in dual M.2 PCIe adapters
superkoning
Indeed. Until that time: buy retail the Raspi Hailo kit, put the M.2 HAT in your drawer (value 10-15 euro?), and use the Hailo M.2 in your system
Mark Tomlin
Will this work to transcribe audio to text? I have like 8kHz audio files where speech to text would be very useful. Perhaps with Whisper AI? Is that possible on this device?
Benjamin
It seems you’re reading in my mind, I was looking for an Hailo 8L to use with a RPi yesterday evening! I can’t wait!
Raspberry Pi Staff Ashley Whittaker
🔮
Michael Kofler
Is it possible to run LLMs on AI Kit? Something like ollama?
Michael S
I’d also be curious about the possibility of running non-computer vision ML models!
petrosilius
Are there somewhere ressources how to train and convert / deploy existing CNN networks for the hailo platform to be used with the RPI software stack? I think this is important for all professional users who want to use their own networks instead of the existing ones. Thanks!
KayS
Are there any recommended add ons to this kit, the Pi camera for example, that would enhance the kit and make it more versatile. I teach an AI unit to HE students and this seems a great way of giving them a hands on experience. Thanks
Charlie Stutzman
Can the object detection setup be used for ai chat gpt like servers as well?
Colin
Will there be an easy way to access Hailo’s dataflow compiler to prepare our own models for use with this?
Nick
DFC is available only for Enterprise Users, so there is no way to convert your existing models being a private person.
Bob Rooney
Can I use a major DB like sql?
colts
I’m looking at Home Assistant + Frigate use. Has anyone compared this with a Coral TPU + M.2 setup?
Luke
Is the Hailio ASIC module sold by itself, or only sold as a bundle with the Raspberry Pi 5 PCIe M.2 HAT+ included?
It would be nice to sell the two components independently, so if you already have the M.2 HAT you don’t need to buy it again.
Justmore5mins
does it support raspberry pi 4b model?
Beyonlo
Is possible to do a fast and secure ( liveness detection) face recognition to use as a solution for the access control? I’m afraid about the Facial spoofing.
Thank you.
Comments are closed