h3-blackwell-runtime/benchmarks/gb10-cute-p1-stream-a-timing-summary.json
2026-08-25 20:30:22 +07:00

50 lines
2.1 KiB
JSON

{
"date": "2026-08-23",
"device": "NVIDIA GB10",
"cutlass_dsl": "4.6.2",
"rows": 128,
"tile_shape_mnk": [128, 128, 128],
"warmup": 5,
"iterations": 20,
"results": {
"attn_qkv_proj": {
"n_tiles": 168,
"k_tiles": 42,
"prepacked_gemm_ms": 0.4769120216369629,
"activation_quantize_ms": 0.05241439938545227,
"reference_complete_projection_ms": 0.49579682350158694,
"streamed_gemm_ms": 4.529531097412109,
"streamed_vs_prepacked": 9.497624073020535,
"streamed_vs_reference_complete": 9.135861471281917,
"repeated_producer_overhead_ms": 4.052619075775146,
"minimum_theoretical_n_reuse_to_cover_quantize_cost": 78
},
"attn_out_proj": {
"n_tiles": 42,
"k_tiles": 56,
"prepacked_gemm_ms": 0.054211199283599854,
"activation_quantize_ms": 0.05055999755859375,
"reference_complete_projection_ms": 0.11322239637374878,
"streamed_gemm_ms": 1.4691216468811035,
"streamed_vs_prepacked": 27.09996580587633,
"streamed_vs_reference_complete": 12.975539238999248,
"repeated_producer_overhead_ms": 1.4149104475975036,
"minimum_theoretical_n_reuse_to_cover_quantize_cost": 28
},
"mlp_fc1": {
"n_tiles": 224,
"k_tiles": 42,
"prepacked_gemm_ms": 0.46705121994018556,
"activation_quantize_ms": 0.051630401611328126,
"reference_complete_projection_ms": 0.45939040184020996,
"streamed_gemm_ms": 5.716193771362304,
"streamed_vs_prepacked": 12.238901275312733,
"streamed_vs_reference_complete": 12.442997825954953,
"repeated_producer_overhead_ms": 5.249142551422119,
"minimum_theoretical_n_reuse_to_cover_quantize_cost": 102
}
},
"producer_overhead_per_n_k_tile_ms_range": [0.0005579445739181673, 0.000601577571257442],
"decision": "Reject one-A-producer-per-output-N-CTA scheduling. Required N reuse factors of 28-102 are not practical with duplicate accumulator state. Prototype a bounded global packed-tile ring or persistent work queue before canonical-shape runtime integration.",
"numerical_status": "All timed streamed outputs remain bit-exact."
}