Community Benchmark Dashboard -- compare ANE performance across Apple Silicon chips #36
Description
Built a community benchmark system and live dashboard to help compare Apple Neural Engine performance across different Apple Silicon chips (M1, M2, M3, M4, M5).
Live Dashboard: https://web-lac-sigma-61.vercel.app
Fork with benchmark tooling: https://github.com/dev-erik/ANE
What it measures:
| Benchmark | What it tells us |
|---|---|
| SRAM probe | ANE SRAM capacity -- where weight spilling starts on your chip |
| Peak TFLOPS | Maximum achievable ANE compute via programmatic MIL |
| Training (CPU cls) | End-to-end training perf with CPU classifier |
| Training (ANE cls) | End-to-end training perf with ANE-offloaded classifier |
How to contribute your results:
# Run the benchmark (takes ~5 min)
bash scripts/run_community_benchmark.sh
# Or skip training if you don't have training data
bash scripts/run_community_benchmark.sh --skip-training
This outputs a JSON file in community_benchmarks/. Submit it via PR or paste it in an issue using the "Benchmark Submission" template.
We'd love to see results from M1, M2, M3, and M5 chips to build a cross-generation comparison of ANE capabilities.
See community_benchmarks/README.md for full details on the JSON schema and submission process.