ane-rooflines / ROOFLINES.md
sbryngelson's picture
Sync roofline datapoints from ANEForge main
aae714a verified
|
Raw
History Blame Contribute Delete
3.8 kB

ANE roofline results

Accumulated across 4 machine(s), 6 submission(s). Each row is a distinct machine (grouped by hardware hash; identical silicon in different chassis stays separate by model identifier).

Add yours: PYTHONPATH=. python3 bench/roofline_suite.py --contributor <your-gh-handle>, commit the JSON it writes under bench/results/rooflines/, regenerate this file with python3 bench/aggregate_rooflines.py, and open a PR. See rooflines/README.md.

Thanks to @axiom-of-choice, @diegobauavi, @sbryngelson for contributing datapoints.

Machines

Chip Model CPU (P+E) GPU Memory macOS Power Code (main merge-base) By Runs
Apple M1 MacBookPro17,1 4+4 8 16 GB unified 26.5.2 (25F84) ac 263e02ff1898 @diegobauavi 1
Apple M1 Max MacBookPro18,2 8+2 32 32 GB unified 26.5.1 (25F80) ac (high-power) 1f8b477a5397 @sbryngelson 2
Apple M2 Pro Mac14,12 6+4 16 32 GB unified 26.5.2 (25F84) ac 64108472d01f (dirty) @axiom-of-choice 2
Apple M5 Pro Mac17,8 6+12 20 48 GB unified 26.5.1 (25F80) ac (high-power) 55a1c6afdad3 (dirty) @sbryngelson 1

Numeric cliffs (correctness rooflines)

Where fp16 math on the engine silently goes wrong (issue #115, docs/cross-chip.md). These are magnitude thresholds, independent of clock, so they are valid even on battery.

Chip Model matmul inf-cliff slice-x16 (Q.4 crop-DMA) reduce exact-sum
Apple M1 MacBookPro17,1 ~32759 clamp @ 4094 <= 2048
Apple M1 Max MacBookPro18,2 ~32759 clamp @ 4094 <= 2048
Apple M2 Pro Mac14,12 ~32759 clamp @ 4094 <= 2048
Apple M5 Pro Mac17,8 ~32759 exact (no clamp) <= 2048

matmul: output magnitude that flips finite->inf (~fp16_max/2). slice: the pre-A16 crop-DMA clamps |value|>4094; 'exact' means the clamp is gone (A16/M5). reduce: largest all-exact integer row-sum (fp16 grid coarsens past 2048).

Performance rooflines (headline)

Peak ANE numbers only; the full per-size sweeps, all-engine comparison, and per-rail watts live in each machine's JSON under perf_rooflines. Perf depends on power/thermals, so the state is shown per row.

Peak fp16 GEMM (TF/s) is the headline compute number -- measured directly on every machine and the most robust cross-chip comparison. Bandwidth/ridge come from the streaming sweep and are more dispatch-overhead-sensitive on smaller/older parts, so read them as indicative. Decode is migrating to a tiled vocab head (#181); see the note under the table.

Chip Model Peak fp16 GEMM (TF/s) Bandwidth (GB/s) Ridge (FLOP/byte) Peak perf/W (GF/s/W) Decode @b1 (tok/s) Power
Apple M1 MacBookPro17,1 2.7 0.86 3149 442 n/a ac
Apple M1 Max MacBookPro18,2 4.6 7.4 616 897 244 ac (high-power)
Apple M2 Pro Mac14,12 3.4 0.98 3429 744 276 ac
Apple M5 Pro Mac17,8 10.0 24 418 918 117 ac (high-power)

n/a decode = a stale datapoint from before the tiled vocab head (#181). The old untiled 32000-vocab lm_head exceeds the 16384 max matmul dimension on the A13-A15 families, so only A16+ (M5) reported. The tiled head fixes this; the older machines re-run to populate it. Until then, use Peak fp16 GEMM (TF/s) as the headline.