same here -- it's also among one of the only things to support Coral devices and RPi video cores.

I would imagine any GPGPU compute-capable pre-CUDA thing probably won't cut it.

I have used Frigate for years, I think early on it didn't support all of those GPUs. So it might be that said videos are out of date.

Maybe my view of frigate and tensorflow (assuming frigate still uses it) is outdated then. I’m referring to tinygrad vs tensorflow when I say GPU support, of course google’s tensorflow is best for google’s TPUs. I’ve had better luck using tinygrad on my personal devices, but I am biased as it’s been a while since I’ve used tensorflow

This would be a good point of differentiation to make on your GitHub page or for a technical audience on your website. Frigate is SOTA in many folks minds, and to show that you are using tinygrad over tensorflow may be a good “modern-ness” signal for that audience.

Edit: another solution in this space shows a list of supported ML runtimes, which would be good info for folks wanting to run on specific hardware. https://github.com/boquila/boquilahub

Supported runtimes list would be nice, but I don't have access to much hardware to test on. I aim to remove most dependencies and support anything that can run tinygrad + ffmpeg