ane-rooflines

Cross-Apple-Silicon performance and fp16-correctness measurements for the Apple Neural Engine (ANE), collected with ANEForge. Each row is one machine (grouped by hardware hash; identical silicon in different chassis stays distinct by model identifier).

See it charted: the ANE leaderboard ranks these machines by peak GEMM, perf-per-watt, and decode throughput.

These are community-contributed submissions mirrored from the public bench/results/rooflines/ in the repo. The full per-size sweeps live in raw/; rooflines.json is the flattened headline table; ROOFLINES.md is the human-readable version.

from datasets import load_dataset
ds = load_dataset("aneforge/ane-rooflines")   # the flattened headline table

Columns

Reading notes

• peak_fp16_gemm_tflops is the most robust cross-chip number. Bandwidth/ridge come from a streaming sweep and are more dispatch-overhead-sensitive on smaller/older parts, so treat them as indicative.
• decode_tok_s currently reports only on A16+ (e.g. M5). The decode benchmark's 32000-vocab head exceeded the 16384 max matmul dimension on the A13-A15 families; a tiled head fixes this and the older machines re-run to populate it. Blank means the run predates the tiled head, not that the chip cannot decode.
• Correctness cliffs are magnitude thresholds, independent of clock, so they are valid even on battery. matmul ~ fp16_max/2 (~32752); slice clamps |value|>4094 on pre-A16 parts and is exact on A16+; reduce is bit-exact for integer sums up to 2048.

Contribute your chip

Run the suite on any Apple Silicon Mac and open a PR:

PYTHONPATH=. python3 bench/roofline_suite.py --contributor <your-gh-handle>

See the roofline drive. More chip generations sharpen the per-family map.

Cite

Bryngelson, S. H. ANEForge: Python for direct computation on the Apple Neural Engine. arXiv:2606.17090 (2026).