// the find
mrdbourke/m1-machine-learning-test
Code for testing various M1 Chip benchmarks with TensorFlow.
A benchmark suite comparing TensorFlow training speed on Apple Silicon (M1/M1 Pro/M1 Max/M2) against Intel Macs, Google Colab, and a Titan RTX, plus a step-by-step guide for getting tensorflow-macos and tensorflow-metal installed. It's for someone who just bought an Apple Silicon Mac and wants to know if it's actually worth training models on, or wants a known-good Conda setup instead of guessing at package versions.
The install walkthrough is unusually thorough for both command-line veterans and people who've never touched conda, with an explicit explanation of what a package manager and an environment even are. Benchmark results are checked into the repo as raw CSVs per device per task, so you can pull the numbers yourself instead of trusting a chart. It covers three genuinely different workload types (CNN on CIFAR10, transfer learning on Food101, a sklearn RandomForest) rather than just one toy MNIST script.
Last pushed April 2024 and pins Python 3.8 with tensorflow-macos/tensorflow-metal versions from the 2.5–2.8 era, so the install steps are stale against current macOS and Python releases. No M3/M4 results at all, which is the hardware most readers landing here now actually own. It's three notebooks and a README, not a library — nothing to import, no CI, no tests, so there's nothing to adopt beyond reading the numbers and maybe copy-pasting the conda commands.