h3-blackwell-runtime/benchmarks/gb10-cute-p2-ring-capacity-summary.json

59 lines
1.9 KiB
JSON
Raw Permalink Normal View History

{
"date": "2026-08-23",
"device": "NVIDIA GB10",
"cutlass_dsl": "4.6.2",
"full_activation_rows": 37810,
"global_scale_policy": "one full-activation Vortex scale reused by every ring chunk",
"attention_output_capacity_sweep": {
"128": {
"ring_bytes": 516096,
"cute_ring_vs_reference": 1.0759724609619812
},
"512": {
"ring_bytes": 2064384,
"cute_ring_vs_reference": 0.93377734918835
},
"1024": {
"ring_bytes": 4128768,
"cute_ring_vs_reference": 0.9562389611635855
},
"2048": {
"ring_bytes": 8257536,
"cute_ring_vs_reference": 0.8562034923214508,
"modeled_canonical_vs_reference": 0.8747956842123643
},
"4096": {
"ring_bytes": 16515072,
"cute_ring_vs_reference": 1.148149277409534
}
},
"selected_row_capacity": 2048,
"selected_capacity_results": {
"attn_qkv_proj": {
"ring_bytes": 6193152,
"chunk_improvement_percent": 10.69608479257264,
"modeled_canonical_improvement_percent": 10.123207712275,
"output_equal": true
},
"attn_out_proj": {
"ring_bytes": 8257536,
"chunk_improvement_percent": 14.37965076785492,
"modeled_canonical_improvement_percent": 12.52043157876357,
"output_equal": true
},
"mlp_fc1": {
"ring_bytes": 6193152,
"chunk_improvement_percent": 11.26063144781996,
"modeled_canonical_improvement_percent": 10.65969035561332,
"output_equal": true
}
},
"policy": {
"attn_qkv_proj": "advance to real-block validation",
"attn_out_proj": "advance to real-block validation",
"mlp_fc1": "advance to real-block validation",
"mlp_fc2": "retain Comfy/cuBLAS fallback"
},
"next_gate": "Execute 2048-row chunks across complete real projections in blocks 0, 24, and 49, including the final partial chunk, and compare total projection/block latency and exact outputs."
}