I was initially testing if TensorFlow was installed correctly on my M1 such that code automatically runs on the GPU outside any context manager. Since we already have runtimes, it would be interesting to compare local results with those of free GPUs in Kaggle.
TLDR: Things look pretty bad for the M1. Not to mention temps were worrying (~97 °C 🌡). This laptop is fanless after all.
Running the benchmarks is easy. After setting up your environment run:
python run.py <compute_env> <benchmark>
# ex: python run.py "Kaggle (P100)" mlp
Current available benchmark names:
mlpvgg
This saves a CSV file in results/<benchmark>/<compute_env>.csv containing the results of the benchmark and automatically updates the plot plots/<benchmark>.png. Note that the plot includes all existing results with the pattern results/<benchmark>/*.csv.
New benchmarks for different architectures and dataset or tasks can be easily created by extending the Benchmark abstract class in benchmarks.py. See existing implementations in that script for the MLP and VGG architectures.
Results for existing benchmarks can be updated by running benchmarks on different environments (e.g. RTX cards, M1 Max, different TF versions):
python run.py <new_compute_env> mlp
- Kaggle P100 and T4 GPU kernels
- Macbook Air 2020 M1 (macOS Monterey v12.1)