When I was a vision engineer at a previous role back in 2022-ish, laptop-class Intel chips used in their SFF NUCs were powerful enough to run small finetuned convolutional nets by dancing across the different on the package: first decode using QSV, and then downsize using the CPU, run the images through an OpenVINO model running on the iGPU, and aggregate results and append metadata using the CPU core. You could have 50% of the hardware in today's high-end security cameras for 10% of the cost.

I have to assume that hardware advancements in the past four years have only pushed the efficiency envelope further.

Some people might want to remove it, but onnx is one of the few ways you can run hw accelerated convolutions

I wonder if it’s going to be some custom encoder/decoder using DNN. I would also want to see some examples!

Well, not that important to me since we already had that in Vapoursynth.