chip stringclasses 4
values | model_identifier stringclasses 4
values | hardware_hash stringclasses 4
values | p_cores int64 4 8 | e_cores int64 2 12 | gpu_cores int64 8 32 | ram_gb int64 16 48 | macos_version stringclasses 2
values | macos_build stringclasses 2
values | aneforge_version stringclasses 4
values | power stringclasses 2
values | contributor stringclasses 3
values | peak_fp16_gemm_tflops float64 2.7 10 | bandwidth_gbps float64 0.86 24 | ridge_flop_per_byte float64 418 3.43k | peak_perf_per_w_gflops float64 442 918 | decode_tok_s float64 117 276 ⌀ | matmul_inf_cliff stringclasses 1
value | slice_x16_cliff stringclasses 2
values | reduce_exact_sum stringclasses 1
value | timestamp_utc timestamp[s]date 2026-08-01 16:39:18 2026-08-05 20:14:25 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Apple M1 | MacBookPro17,1 | dcf5b8a7dd09 | 4 | 4 | 8 | 16 | 26.5.2 | 25F84 | 0.2.1.dev42 | ac | diegobauavi | 2.7 | 0.86 | 3,149 | 442 | null | ~32759 | clamp @ 4094 | <= 2048 | 2026-08-01T20:23:15 |
Apple M1 Max | MacBookPro18,2 | 8dc9588b5db6 | 8 | 2 | 32 | 32 | 26.5.1 | 25F80 | 0.2.0 | ac (high-power) | sbryngelson | 4.59 | 7.44 | 616 | 897 | 244 | ~32759 | clamp @ 4094 | <= 2048 | 2026-08-05T20:14:25 |
Apple M2 Pro | Mac14,12 | 723396ad9091 | 6 | 4 | 16 | 32 | 26.5.2 | 25F84 | 0.2.1.dev33 | ac | axiom-of-choice | 3.36 | 0.98 | 3,429 | 744 | 276 | ~32759 | clamp @ 4094 | <= 2048 | 2026-08-05T17:34:43 |
Apple M5 Pro | Mac17,8 | c38210cfc8ea | 6 | 12 | 20 | 48 | 26.5.1 | 25F80 | 0.1.4.dev32+gb8dc90fe6.d20260624 | ac (high-power) | sbryngelson | 10.04 | 24 | 418 | 918 | 117 | ~32759 | exact (no clamp) | <= 2048 | 2026-08-01T16:39:18 |
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
| column | meaning |
|---|---|
chip, model_identifier, hardware_hash |
machine identity |
p_cores, e_cores, gpu_cores, ram_gb |
CPU perf/efficiency cores, GPU cores, unified memory |
macos_version, macos_build, aneforge_version |
software the run was recorded under |
power |
ac, ac (high-power), or battery at run time |
contributor |
GitHub handle who submitted the run |
peak_fp16_gemm_tflops |
headline compute peak (measured on every machine) |
bandwidth_gbps, ridge_flop_per_byte |
streaming bandwidth and the ridge point |
peak_perf_per_w_gflops |
peak GFLOP/s per watt |
decode_tok_s |
single-stream LLM decode throughput |
matmul_inf_cliff, slice_x16_cliff, reduce_exact_sum |
fp16 correctness cliffs (magnitudes where the engine silently returns a wrong answer) |
timestamp_utc |
when the run was recorded |
Reading notes
peak_fp16_gemm_tflopsis 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_scurrently 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);sliceclamps|value|>4094on pre-A16 parts and is exact on A16+;reduceis 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).
- Downloads last month
- -