mirror of
https://github.com/luxfi/aivm.git
synced 2026-10-30 17:02:34 +00:00
zeekay
ea1f33857b
Splits AttestationApply into two Metal kernels — single-thread canonical-order locate (preserves byte-equal slot assignment with the CPU oracle) followed by a parallel writeback dispatched at threadgroup width 256, one thread per slot. Same arch hook in place for ProvenanceApply via op_slot tracking; the apply itself stays single-pass because UpdateWeights' version counter is order- dependent. AnchorApply stays single-pass (parent_root chain integrity is fundamentally sequential). EpochTransition stays single-thread on Metal — multi-dispatch parallel split was tried, dispatch latency on the M1 Max integrated GPU exceeded the saved keccak work at every workload size. Determinism: 47 / 47 tests pass. New size sweep covers small / medium / large workloads byte-equal CPU vs Metal vs WGSL (XVM caught a size-dep race in its parallel rewrite; AIVM's locate-then-writeback is byte-identical at all sizes by construction, but the sweep is cheap insurance). Benchmark: FullRound Metal essentially flat vs v0.58.3 on the M1 Max integrated GPU — the architectural change is live but per-dispatch latency (~1-2 ms) on the integrated GPU swamps the parallel writeback savings at every workload size. The split is correct and ready for hardware where dispatch latency is amortised differently (Linux+CUDA, dedicated cards).
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 1.363 1.358 1.397 1.403 73370 1.00x 10
|
|
cpu medium 4.535 4.531 4.610 4.612 220524 1.00x 10
|
|
cpu large 14.492 14.558 14.654 14.671 690036 1.00x 10
|
|
cpu xlarge 40.694 40.780 41.230 41.359 2457378 1.00x 10
|
|
metal small 26.513 26.490 26.629 26.644 3772 0.05x 10
|
|
metal medium 83.533 83.529 83.588 83.614 11971 0.05x 10
|
|
metal large 260.996 261.246 261.517 261.566 38315 0.06x 10
|
|
metal xlarge 706.062 706.084 706.333 706.354 141631 0.06x 10
|
|
wgsl small 2029.272 1338.519 6980.984 10319.810 49 0.00x 10
|
|
wgsl medium 1711.537 1840.082 2338.952 2663.783 584 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 1.217 1.208 1.264 1.266 82185 1.00x 10
|
|
cpu medium 3.162 3.178 3.236 3.250 316213 1.00x 10
|
|
cpu large 3.232 3.239 3.319 3.329 3093605 1.00x 10
|
|
cpu xlarge 4.259 4.252 4.387 4.395 23481528 1.00x 10
|
|
metal small 737.528 61.214 2755.214 3726.970 136 0.00x 10
|
|
metal medium 87.174 89.383 95.469 96.318 11471 0.04x 10
|
|
metal large 1156.552 795.390 3800.330 5032.623 8646 0.00x 10
|
|
metal xlarge 180.505 118.280 379.868 405.825 554001 0.02x 10
|
|
wgsl small 192.470 193.183 193.929 193.963 520 0.01x 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 0.075 0.075 0.075 0.075 67013 1.00x 10
|
|
cpu medium 0.629 0.629 0.635 0.635 79486 1.00x 10
|
|
cpu large 3.095 3.096 3.174 3.175 161526 1.00x 10
|
|
cpu xlarge 3.195 3.207 3.222 3.224 1565153 1.00x 10
|
|
metal small 116.876 24.694 515.694 686.389 43 0.00x 10
|
|
metal medium 52.624 45.178 97.569 101.517 950 0.01x 10
|
|
metal large 173.517 86.548 545.425 616.901 2882 0.02x 10
|
|
metal xlarge 308.984 296.086 654.892 750.343 16182 0.01x 10
|
|
wgsl small 15.203 15.164 15.437 15.441 329 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 0.091 0.091 0.091 0.091 109945 1.00x 10
|
|
cpu medium 0.811 0.810 0.830 0.831 123258 1.00x 10
|
|
cpu large 8.038 8.034 8.239 8.264 124404 1.00x 10
|
|
cpu xlarge 33.040 33.032 33.402 33.447 302667 1.00x 10
|
|
metal small 213.203 239.249 400.742 429.184 47 0.00x 10
|
|
metal medium 253.513 69.160 886.682 1195.689 394 0.00x 10
|
|
metal large 629.953 133.506 1378.933 1379.123 1587 0.01x 10
|
|
metal xlarge 900.272 528.376 1772.237 1773.299 11108 0.04x 10
|
|
wgsl small 6914.505 1263.067 28113.204 36233.111 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
|
|
|
|
=== EpochTransition - single-mode kernel breakdown ===
|
|
backend size mean(ms) p50(ms) p95(ms) p99(ms) ops/sec vs CPU iters
|
|
cpu small 0.014 0.014 0.018 0.020 6918548 1.00x 10
|
|
cpu medium 0.013 0.013 0.014 0.014 74142175 1.00x 10
|
|
cpu large 0.014 0.013 0.016 0.018 723102399 1.00x 10
|
|
cpu xlarge 0.013 0.013 0.013 0.013 7741975442 1.00x 10
|
|
metal small 6768.602 4113.548 15704.264 17221.962 15 0.00x 4*
|
|
metal medium 3.271 2.786 5.671 6.105 305754 0.00x 10
|
|
metal large 2.613 2.470 3.248 3.561 3827135 0.01x 10
|
|
metal xlarge 2.558 2.564 2.641 2.682 39091069 0.01x 10
|
|
wgsl small 4.045 4.061 4.134 4.147 24724 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
|