h3-blackwell-runtime/benchmarks/gb10-cute-p2-ring-summary.json
2026-08-25 20:30:22 +07:00

49 lines
1.6 KiB
JSON

{
"date": "2026-08-23",
"device": "NVIDIA GB10",
"cutlass_dsl": "4.6.2",
"row_capacity": 128,
"warmup": 10,
"iterations": 100,
"producer": "vortex_native_quantize_nvfp4_into",
"consumer": "CuTe SM121 prepacked block-scaled GEMM",
"results": {
"attn_qkv_proj": {
"ring_bytes": 387072,
"reference_ms": 0.47588897705078126,
"ring_ms": 0.4637001419067383,
"improvement_percent": 2.56127704818477,
"fp4_difference_count": 0,
"block_scale_difference_count": 0,
"output_equal": true,
"decision": "experimental; require full-M chunk-size validation"
},
"attn_out_proj": {
"ring_bytes": 516096,
"reference_ms": 0.10898783683776855,
"ring_ms": 0.09652607917785644,
"improvement_percent": 11.43408110618966,
"fp4_difference_count": 0,
"block_scale_difference_count": 0,
"output_equal": true,
"decision": "advance to chunk-size and real-block validation"
},
"mlp_fc1": {
"ring_bytes": 387072,
"reference_ms": 0.5960015869140625,
"ring_ms": 0.6573414611816406,
"improvement_percent": -10.29189780939672,
"fp4_difference_count": 0,
"block_scale_difference_count": 0,
"output_equal": true,
"decision": "reject current ring consumer; retain reference path"
}
},
"policy": {
"attn_out_proj": "continue",
"attn_qkv_proj": "experimental",
"mlp_fc1": "fallback",
"mlp_fc2": "fallback"
},
"next_gate": "Sweep ring row capacity using one global activation scale, then validate the accepted attention-output path in real blocks 0, 24, and 49."
}