Couldn’t they reuse the tensor cores that are shipped in every device at this point? There are already lots of papers on compressing images using deep learning, I don’t see any reason why the companies couldn’t make a video standard that relies on that hardware.
having a hardware encoder and decoder on a device is super useful for streaming content of that device. Not sure I would want to use other compute for that, that compute is much better used doing CV on the video stream :)
Why do you think so? Those tensor processors are actually already optimized for video processing: all of the complex postprocessing in the iPhone camera app is done by the tensor cores inside the M1 chip. I wouldn't be suprised if it would already far be able outperform the mentioned codecs, but of course it needs lots of software development that can only be done by the big companies.
A codec it’s static, almost not changing at all over a decade. This allow you to implement it as a single purpose hardware which is orders of magnitude more efficient and fast than code running in a multipurpose chip, tensor or not.
For things that evolve fast, as deep learning, an programmable chip is the right choice.
The iPhone doesn't yet use M1. Besides, post-processing a video is one thing, encoding is completely different. What Apple does with the neural processing is most likely the analysis of the content, not the "editing".
In something like a mobile device, every watt counts. If it takes more energy to decode video on the tensor cores than it does to have a dedicated hardware block, you keep the hardware video decoder.
