OpenTube

Raspberry Pi AI HAT+ 2 Reality Check

Hardware.ai

12,914 views

Is the Raspberry Pi AI HAT+ 2 actually an AI accelerator? In this video I run real benchmarks and test Hailo 10H performance on Raspberry Pi 5, including time-to-first-token, VLM inference, and large prompt prefills. I verify the official claims with practical tests and share if this GenAI HAT makes sense — and when.

Test scripts:
https://gist.github.com/AIWintermuteA...
Official Pi AI Hat+ 2 page:
https://www.raspberrypi.com/products/...

Some alternative option (RK3576)
https://github.com/airockchip/rknn-ll...
Finally! An SBC for LLMs
Waveshare Luckfox Core3576 Edge Computing Development Board,Rockchip RK3576 Octa-Core 2.2GHz,6 TOPS Computing Power NPU
https://s.click.aliexpress.com/e/_EJR...

Sources:
https://hailo.ai/blog/bringing-genera...
https://hailo.ai/products/hailo-softw...
https://github.com/hailo-ai/hailo-app...
https://hailo.ai/developer-zone/docum...
https://www.raspberrypi.com/news/when...
https://www.cnx-software.com/2026/01/...
AI this, AI that, Here's Raspberry Pi's ne...

5 Comments

@geoffhubbard7815 1 month ago

I had understood the AI Hat2 not to be "faster" per se, but only that it offloaded the effort to the AI hat so you could still run heavier apps on the pi at the same time
I bought one, but the tokens per second bothers me more than I imagined it would... I am going to run a recognition engine and a robotic arm to sort my HUGE collection of MTG cards as a next test to hopefully salvage what i am afraid was a bad buy.

@majorxbla9432 6 months ago

Ahrg their advertisement got me, and I already bought the GenAIHat. This video came to late for me. But I've read in some forum posts that rather than using GenAIHat for GenAI and LLMs it is quiet suitable for good old multi layer perceptron tasks. Performance in object recognition or on models like whisper for STT would be interesting. Thx for that critical review.

@AntonMaltsev 6 months ago

The interesting question is whether it is because of Hailo or because of a single PCIe line available on RPi.
But, as far as I understand, for AI HAT+ 2, it's only one line available?

@malcolmhutchison 2 months ago

Which Pi5 verfion did you use? (the two htop screens suggest a 4Gb and a 2Gb model, which would be even more damning)

@onekoneb 1 month ago

I got mine up and running and I tell ya, getting results from questions takes a bit of time. Some models are faster than others, but it's much faster and easier using Google AI search than this.