Likewise, I would like to know how I can use it to build and compile my own models. These would not be vision related. I want to know how I can do low level programming of the device so that I could continue the pattern recognition research I was doing thirty years ago using a Sun Sparcstation 1+. As a former accredited expert in pattern recognition research, I find that everything is now obfuscated by the most appalling level of jargon.The RPi5 is vastly more powerful than the Sparcstation 1+ I used to use more than thirty years ago, and with this Hailo 8 device, I should be able to have the power of one of the Cray Supercomputers I used to drool over. But to use it the way I want to, I need to be able to design and compile my systems on my Raspberry Pi because I long since gave up with other computers, and only use Raspberry Pis now. I am a little old lady who first learnt to program in the ’60s, and who is now long since retired. I do not want to have to build an expensive x86 machine just to build and compile models, and would rather not have to learn the appalling mass of jargon which has proliferated since I retired.

Does it supports photon vision

Dear Jon,
To obtain suitable cooling, I would recommend housing your RPi5 in the new Bumper, and using the active cooler underneath the hat. Using the Bumper with the Active Cooler will then also permit you to fit the official Hailo cooling block to the mounting holes provided on the AI Hat+. This combination will give you a neat compact assemblage with the best cooling.

I really hope so!

At the rate that the Raspberry Pi team have been developing new technology, I wouldn’t be suprised if the Pi7 has a tiny integrated inflatable monitor which at the touch of a button expands to become a 32 inch touch-screen.

I have used an dual nvme m.2 Hat. With m.2 disk and Hailo-8L, but yes the share the same pcie-port … But the write to disk performance is much better than SD-card

Yes there are PCIe multiplexes. But make sure the one you but has external power connection, so the power the rpi outputs if enough for both of them. Also this configuration is not checked, so please share your results (:

You have hit the nail on the head here. The RPi5 is by far the best and most versatile computer I have used in nearly 60 years of computing; but the AI hats are the opposite. I am completely unable to use my standard AI Hat to even repeat the pattern recognition research I published 30 years ago!

The Raspberry Pi community is made up of a wonderful polyglot of people with all sorts of levels of experience and skills, with the common factor being strong creative imaginations along with a willingness to experiment in all sorts of ways with their Pis. It would be in Hailo’s interest to open up more to the Raspberry Pi community so that we can do all the amazingly inventive things with their products that we do with our Pis. There would then be all the astounding designs, crazy attempts, and sheer potty gimmicks that make up the path of natural technological development. So a big PLEASE to Eben Upjohn to persuade Hailo to open up to us.

Specialized video AI is fine for me.
But it is quiet odd that they removed hardware support for their cameras and common web cams, did not replaced it with working software and show up with video AI.

LLMs require quite a bit more RAM than is available on the HAILO 8 devices and on the Pi itself.
There is a new device coming, the HAILO 10L that will extend low cost, low power devices beyond vision tasks. It has the required RAM onboard. It is M.2 form factor but I’ve no idea if it will work with Raspberry PI as it is Gen3 PCIe with 4 lanes. It would require the software component to be developed if it were possible use it on one lane.

Dear Mark,
Processing of Speech-to-text, textual data, audio data, and numerical data, are all common fields for application of machine learning. All should be supported. From what little I can gather from the apparent architecture of the Hailo device, it would appear to use matrix operations which are integer only. This might limit the efficacy of any audio processing needing to use Fourier Analysis, or using Fourier Transforms. Likewise, the integer-only operations would limit the numerical data types which could be used with this device.

Dear Anders,
Wait till you get to my age. Then you will find you make far more of these little errors. Errors like this which I have made on this discussion, and subsequently spotted, have included; leaving the “r” out of “matrices” to read “matices”, and unfortunately calling Eben Upton “Eben Upjohn”! I call these “Senior Moments”.

Dear Givi,
The product brief for the AI Hat+ does not give its power requirements. The current requirement is going to be very much larger than for an SSD. Couple this fact to the AI Hat+ automatically switching the PCIe bus to PCIe3, which will probably increase the current draw of the SSD appreciably, and you should be able to see that this combination is quite likely to go over the Pi5 power budget. I wouldn’t like to guess what problems any such current shortfall would cause. I can only say that I wouldn’t attempt it myself. A better option would be to use two Pi5 computers, one for the workstation with SSD, and the other as a headless server with the AI Hat+. I like to err on the side of caution, though I accept you might be more willing to live dangerously. ;-)