File size: 3,298 Bytes
48aae68 9f4c27e 48aae68 95c0f7c 48aae68 9f4c27e 48aae68 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | ---
license: mit
pretty_name: ANE Rooflines
tags:
- benchmark
- apple-neural-engine
- apple-silicon
- roofline
- hardware
size_categories:
- n<1K
configs:
- config_name: default
data_files: rooflines.json
---
# ANE Rooflines
Cross-Apple-Silicon performance and fp16-correctness measurements for the Apple
Neural Engine (ANE), collected with [ANEForge](https://github.com/sbryngelson/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](https://huggingface.co/spaces/aneforge/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/`](https://github.com/sbryngelson/ANEForge/tree/main/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.
```python
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_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:
```sh
PYTHONPATH=. python3 bench/roofline_suite.py --contributor <your-gh-handle>
```
See the [roofline drive](https://github.com/sbryngelson/ANEForge/issues/137). 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).
|