mirror of
https://github.com/luxfi/aivm.git
synced 2026-10-30 17:02:34 +00:00
zeekay
b603d5c184
The kernels are correctness-first (4-way byte-equivalent CPU/Metal/CUDA/WGSL) but dispatch threadgroups of (1,1,1) — single-thread GPU shaders. CPU peak is 24.4M attestations/sec (xlarge AttestationApply), Metal peak is 1.37M. GPU never wins on this workload shape on this hardware. The substrate's value is offload-from-validator-CPU during consensus, not raw throughput. v0.59+ should parallelise the per-op loop within each phase to lift the per-thread N from 1 to ~1024. Adds: - test/aivm_benchmark.cpp: CPU vs Metal vs WGSL across small/medium/large/ xlarge for FullRound + 4 single-mode kernel breakdowns (3 warmup, 10 measured iters, 60s per-cell budget cap) - CMakeLists.txt: aivm-benchmark target at -O3 -DNDEBUG - BENCHMARKS.md, BENCHMARKS_RAW.txt: full results, methodology, hardware
82 lines
6.6 KiB
Plaintext
82 lines
6.6 KiB
Plaintext
[aivm_benchmark] starting (warmups=3, iters=10, cap=60s)
|
|
backend: metal -> Apple M1 Max
|
|
backend: wgsl -> wgpu-native
|
|
backend: cpu -> reference oracle
|
|
|
|
=== FullRound (Attestation + Provenance + Anchor + EpochTransition) ===
|
|
backend size mean(ms) p50(ms) p95(ms) p99(ms) ops/sec vs CPU iters
|
|
cpu small 1.337 1.338 1.381 1.390 74781 1.00x 10
|
|
cpu medium 4.414 4.417 4.511 4.538 226549 1.00x 10
|
|
cpu large 13.927 13.937 14.071 14.075 718053 1.00x 10
|
|
cpu xlarge 39.363 39.329 39.934 39.937 2540461 1.00x 10
|
|
metal small 26.303 26.279 26.425 26.427 3802 0.05x 10
|
|
metal medium 83.325 83.325 83.491 83.528 12001 0.05x 10
|
|
metal large 261.286 261.335 261.533 261.560 38272 0.05x 10
|
|
metal xlarge 695.354 695.129 697.044 697.177 143812 0.06x 10
|
|
wgsl small 3832.203 1711.237 10223.456 10997.248 26 0.00x 10
|
|
wgsl medium 4192.341 3085.777 9928.841 12765.474 239 0.00x 10
|
|
wgsl large skip skip skip skip — — 0
|
|
wgsl xlarge skip skip skip skip — — 0
|
|
|
|
=== AttestationApply - single-mode kernel breakdown ===
|
|
backend size mean(ms) p50(ms) p95(ms) p99(ms) ops/sec vs CPU iters
|
|
cpu small 1.186 1.178 1.222 1.229 84344 1.00x 10
|
|
cpu medium 3.434 3.336 3.932 3.950 291241 1.00x 10
|
|
cpu large 3.191 3.163 3.358 3.388 3134125 1.00x 10
|
|
cpu xlarge 4.091 4.078 4.170 4.202 24440989 1.00x 10
|
|
metal small 23.370 23.371 23.428 23.457 4279 0.05x 10
|
|
metal medium 58.530 58.539 58.622 58.645 17085 0.06x 10
|
|
metal large 72.738 72.760 72.924 72.967 137480 0.04x 10
|
|
metal xlarge 72.772 72.807 72.896 72.923 1374151 0.06x 10
|
|
wgsl small 4749.315 4955.738 8403.558 8852.765 21 0.00x 10
|
|
wgsl medium skip skip skip skip — — 0
|
|
wgsl large skip skip skip skip — — 0
|
|
wgsl xlarge skip skip skip skip — — 0
|
|
|
|
=== ProvenanceApply - single-mode kernel breakdown ===
|
|
backend size mean(ms) p50(ms) p95(ms) p99(ms) ops/sec vs CPU iters
|
|
cpu small 0.071 0.071 0.073 0.074 70381 1.00x 10
|
|
cpu medium 0.639 0.652 0.657 0.659 78300 1.00x 10
|
|
cpu large 3.823 3.896 3.905 3.908 130774 1.00x 10
|
|
cpu xlarge 3.144 3.101 3.338 3.340 1590413 1.00x 10
|
|
metal small 3.573 3.557 3.778 3.831 1399 0.02x 10
|
|
metal medium 12.922 12.914 12.990 12.997 3869 0.05x 10
|
|
metal large 56.848 55.743 61.912 65.818 8795 0.07x 10
|
|
metal xlarge 73.731 73.741 73.847 73.861 67814 0.04x 10
|
|
wgsl small 3586.027 1427.623 9267.936 9949.879 1 0.00x 10
|
|
wgsl medium skip skip skip skip — — 0
|
|
wgsl large skip skip skip skip — — 0
|
|
wgsl xlarge skip skip skip skip — — 0
|
|
|
|
=== AnchorApply - single-mode kernel breakdown ===
|
|
backend size mean(ms) p50(ms) p95(ms) p99(ms) ops/sec vs CPU iters
|
|
cpu small 0.118 0.117 0.119 0.120 84959 1.00x 10
|
|
cpu medium 0.959 0.941 1.025 1.035 104237 1.00x 10
|
|
cpu large 7.774 7.768 7.798 7.800 128631 1.00x 10
|
|
cpu xlarge 32.268 32.135 32.789 32.794 309902 1.00x 10
|
|
metal small 3.854 3.856 4.005 4.031 2595 0.03x 10
|
|
metal medium 17.027 17.025 17.103 17.128 5873 0.06x 10
|
|
metal large 137.713 136.936 139.982 139.984 7261 0.06x 10
|
|
metal xlarge 551.869 551.558 553.777 555.067 18120 0.06x 10
|
|
wgsl small 1609.601 511.818 4736.407 4759.745 6 0.00x 10
|
|
wgsl medium skip skip skip skip — — 0
|
|
wgsl large skip skip skip skip — — 0
|
|
wgsl xlarge skip skip skip skip — — 0
|
|
|
|
=== EpochTransition - single-mode kernel breakdown ===
|
|
backend size mean(ms) p50(ms) p95(ms) p99(ms) ops/sec vs CPU iters
|
|
cpu small 0.016 0.016 0.017 0.017 6075925 1.00x 10
|
|
cpu medium 0.016 0.016 0.017 0.017 60759248 1.00x 10
|
|
cpu large 0.016 0.016 0.016 0.016 608987437 1.00x 10
|
|
cpu xlarge 0.016 0.016 0.016 0.016 6082096134 1.00x 10
|
|
metal small 2.267 2.261 2.309 2.309 44116 0.01x 10
|
|
metal medium 2.318 2.319 2.350 2.354 431344 0.01x 10
|
|
metal large 1690.897 2.525 8803.212 12812.606 5914 0.00x 10
|
|
metal xlarge 2.483 2.423 2.784 2.895 40269539 0.01x 10
|
|
wgsl small 1174.440 321.541 5212.714 7972.537 85 0.00x 10
|
|
wgsl medium skip skip skip skip — — 0
|
|
wgsl large skip skip skip skip — — 0
|
|
wgsl xlarge skip skip skip skip — — 0
|
|
|
|
[aivm_benchmark] done
|