tcapelle / apple_m1_pro_python

A collection of ML scripts to test the M1 Pro MacBook Pro

Apple Silicon DL benchmarks

Currently we have PyTorch and Tensorflow that have Metal backend.

Results

Varied results across frameworks:

• Apple M1Pro Pytorch Training Results
• Apple M1Pro Tensorflow Training Results

Tensorflow Resnet50:

PyTorch Resnet50:

• Difference between CPU and GPU
• Comparing with Nvidia

PyTorch Bert

• Running a Bert from Huggingface

Pytorch

We have official PyTorch support! check pytorch folder to start running your benchmarks

Tensorflow

You can run tensorflow benchmarks by going to the tensorflow folder.

apple_m1_pro_python's Issues

Missing one dependency `datasets`

Presumably it's using huggingface datasets

on batchsize in the wandb experiments

hey - great work, i just have a question regarding the batchsizes chosen for the benchmark experiment you shared on wandb

for example in the first table, if i expand all runs, i can see that it says

• 1080 ti uses batchsize 160
• A5000 uses batchsize 128
• ...
• M1Pro uses batchsize 96
• M1_7 uses batchsize 128
• M1Max uses batchsize 128

and this inconsistency persists throughout all tables in the benchmark, for example if you look at the resnet graph, and expand the train_full_model tab, it says batchsize-wise M1Pro uses 48, RTX 3090 uses 512, and M1Max uses 128, which are wildly inconsistent.

is this just a typo in the table? or is there something wrong with the underlying experiment too? if this indeed indicates diff batchsizes across different benchmarks, it will affect the performance obviously no?

Update to link needed for Tensorflow documentation

tf_apple.yml is now in scripts folder. https://github.com/tcapelle/apple_m1_pro_python/blob/main/tensorflow/scripts/tf_apple.yml

It needs to be updated in the documentation here.
https://github.com/tcapelle/apple_m1_pro_python/tree/main/tensorflow

I made the PR #7

Thank you for your work.