Thread

I did a couple simple benchmarks for tensorflow on a M1 Max. TL;DR: M1 GPU is only 60% slower than a GTX1660S, but the CPU is 4x faster than the AMD. I took this simple benchmark that trains a ResNet5 on the cifar data set. medium.com/analytics-vidh…

For comparison, I used a Linux box with an 8 core AMD CPU and a seasoned GTX 1660S. One epoch took 61s on the GPU, 1376s on the CPU.

I followed these docs to install tensorflow with M1 support from Apple developer.apple.com/metal/tensorfl… Results: one epoch on the M1 GPU took 96s, and 410s on the (10 core) CPU.

So in summary, the M1 Max GPU is about 57% slower than the GTX, but the CPU is 4x faster than the AMD. And the MBP doesn‘t even get hot. You can barely hear the fan.

I also ran it on Google‘s colab. I got a K80 and one epoch took 162s. Pretty nice.