merlinpan

Someone have benchmarked M1 on tensorflow: https://github.com/tlkh/tf-metal-experiments/blob/main/README.md#experiments-and-benchmarks So on the M1Max (macbook pro), they reached ~40W at ~1/8 the performance of nvidia 3090 (at 1/8 the power draw approximately as...

I suspect it wouldn't give better performance than opencl, except for rare platforms that have AMD's tensor support (server processors only?)

One question there is whether katago's weights are amenable to lower precision or int8. Because that seems to be what the edge tpu supports.