Community Benchmark Dashboard -- compare ANE performance across Apple Silicon chips

[dev-erik]

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.

[xamgore]

M3 Max here. Where can I find the training data?

UPD. Ah, I see. https://github.com/maderix/ANE?tab=readme-ov-file#training-data