XuyangShen/m1_pytorch_performance

Mac M1 Native PyTorch

• GitHub Notion

---

Performance Experiment Results

CIFAR 10 ResNet50	M1 Pro 8 Core	M1 Pro 14 GCore	NV 3070 Laptop	NV A100-40g x1	NV A100-40g x8	NV A100-40g x16
Training Time (seconds/epoch)	240.17	104.70 (2.3x)	7.82	4.50	2.22	2.01
Test Acc (% after 20 epochs)	45.88	43.38	56.11	43.40	31.02	31.02

• Environment
  • Code: Train a ResNet-50 classifier on CIFAR10
  • M1 Pro: PyTorch 1.12, Python 3.10
  • 3070 Laptop: cuda 11.6, PyTorch 1.11, Python 3.10
  • A100-40g: cuda 11.2, PyTorch1.11, Python 3.10, Distributed Data Parallel (DDP)
  • All results are averaged from 3 runs.

Install Conda

• Check my video: https://www.bilibili.com/video/BV1cR4y1w7CF/

• Miniconda: https://docs.conda.io/en/latest/miniconda.html

  Miniconda3 macOS Apple M1 64-bit pkg

• Miniforge: https://github.com/conda-forge/miniforge

  Miniforge3-MacOSX-arm64

Verify conda platform is ARM

$ conda info

...
platform : osx-arm64
...

Create new environment and activate

$ conda create -n torch1.12 python=3.10 #name=torch1.12 with python3.10 
$ codna activate torch1.12 

Install PyTorch1.12

• https://pytorch.org/get-started/locally/ (Preview - Mac - Pip)

$ conda install pytorch torchvision -c pytorch-nightly

• Install other packages

$ pip install notebook