I see no mention of simd in these threads. The only thing that has made high data throughput possible is vector operations. What's the performance of libjpegturbo or hevc in software?
Apple CPUs have great NEON performance, so if those libraries have NEON paths, they'll scream.
It supports NEON like most phone CPUs. The new thing for a desktop is sharing memory for the CPU and GPU with a much faster bus instead of having the GPU sit on a slower bus with its own fast memory, so if your workload can benefit from using the GPU, you don't have to worry about host to device transfer.
I can't find a number for it on M1 but I'm guessing a memory to SIMD register move is still a lot faster than faster CPU-GPU shared buffers, i.e. Even if it is a powerful tool it won't scale in the same way.