95 lines
3.7 KiB
JSON
95 lines
3.7 KiB
JSON
{
|
|
"status": "complete",
|
|
"device": "NVIDIA GB10",
|
|
"compute_capability": "SM121",
|
|
"workload": {
|
|
"resolution": [1344, 768],
|
|
"frames": 124,
|
|
"tokens": 37810,
|
|
"block": 24,
|
|
"seed": 440420
|
|
},
|
|
"method": {
|
|
"flops": "2*M*N*K dense-equivalent operations",
|
|
"l2_traffic": "Nsight Compute lts__t_bytes.sum with cache control disabled",
|
|
"off_chip_proxy": "32 bytes times warm-cache L2 read plus write lookup-miss sectors",
|
|
"practical_compute_ceiling_tflops_per_s": 321.455895,
|
|
"theoretical_memory_bandwidth_gb_per_s": 273.0,
|
|
"practical_crossover_flops_per_byte": 1177.49412,
|
|
"note": "GB10 exposes no direct physical DRAM-byte counter; classification uses the documented bandwidth and the warm-cache off-chip request proxy."
|
|
},
|
|
"gemms": [
|
|
{
|
|
"role": "attention_qkv",
|
|
"shape_mnk": [37810, 21504, 5376],
|
|
"kernel": "standard",
|
|
"duration_ms": 27.195296,
|
|
"dense_equivalent_tflops": 8.742088,
|
|
"achieved_tflops_per_s": 321.455895,
|
|
"l2_requested_gb": 40.129071,
|
|
"off_chip_proxy_gb": 4.368326,
|
|
"off_chip_arithmetic_intensity": 2001.244307,
|
|
"ideal_unique_bytes_gb": 1.805498,
|
|
"off_chip_amplification": 2.419458,
|
|
"classification": "compute-bound"
|
|
},
|
|
{
|
|
"role": "attention_output",
|
|
"shape_mnk": [37810, 5376, 7168],
|
|
"kernel": "standard",
|
|
"duration_ms": 9.25072,
|
|
"dense_equivalent_tflops": 2.914029,
|
|
"achieved_tflops_per_s": 315.005686,
|
|
"l2_requested_gb": 13.24078,
|
|
"off_chip_proxy_gb": 1.337147,
|
|
"off_chip_arithmetic_intensity": 2179.28899,
|
|
"ideal_unique_bytes_gb": 0.580659,
|
|
"off_chip_amplification": 2.302809,
|
|
"classification": "compute-bound"
|
|
},
|
|
{
|
|
"role": "mlp_fc1",
|
|
"shape_mnk": [37810, 28672, 5376],
|
|
"kernel": "standard",
|
|
"duration_ms": 37.024512,
|
|
"dense_equivalent_tflops": 11.656118,
|
|
"achieved_tflops_per_s": 314.821641,
|
|
"l2_requested_gb": 53.507764,
|
|
"off_chip_proxy_gb": 5.823759,
|
|
"off_chip_arithmetic_intensity": 2001.476751,
|
|
"ideal_unique_bytes_gb": 2.369218,
|
|
"off_chip_amplification": 2.458093,
|
|
"classification": "compute-bound"
|
|
},
|
|
{
|
|
"role": "mlp_fc2",
|
|
"shape_mnk": [37810, 5376, 14336],
|
|
"kernel": "stream-k",
|
|
"duration_ms": 54.973312,
|
|
"dense_equivalent_tflops": 5.828059,
|
|
"achieved_tflops_per_s": 106.016148,
|
|
"l2_requested_gb": 26.07999,
|
|
"off_chip_proxy_gb": 12.161131,
|
|
"off_chip_arithmetic_intensity": 479.236573,
|
|
"ideal_unique_bytes_gb": 0.754785,
|
|
"off_chip_amplification": 16.112046,
|
|
"classification": "bandwidth-bound",
|
|
"roofline_at_273_gb_per_s_ms": 44.546268,
|
|
"measured_bandwidth_efficiency_percent": 81.032534,
|
|
"traffic_required_to_reach_practical_compute_ridge_gb": 4.949544,
|
|
"required_traffic_reduction_percent": 59.299,
|
|
"practical_compute_floor_ms": 18.130197
|
|
}
|
|
],
|
|
"decision": {
|
|
"primary_target": "mlp_fc2",
|
|
"do_not_target_first": ["attention_qkv", "attention_output", "mlp_fc1"],
|
|
"reason": "QKV, attention output, and FC1 cluster at 315-321 dense-equivalent TFLOP/s and lie above the practical ridge point. FC2 uses stream-K, reaches only 106 TFLOP/s, and generates 12.16 GB of off-chip request traffic, 16.1x its ideal unique bytes.",
|
|
"next_experiment": "Benchmark an exact FC2 schedule that reduces K-split/stream-K rereads and reuses the fused SwiGLU-produced activation tiles. Require bit-exact BF16 output and at least 3% complete-block improvement before integration."
|
|
},
|
|
"source_artifacts": [
|
|
"benchmarks/gb10-fully-fused-block24-fresh-l2.csv",
|
|
"benchmarks/gb10-fully-fused-block24-fresh-l2-misses-warm.csv",
|
|
"benchmarks/gb10-fully-fused-fresh-nsight-summary.json"
|
|
]
|
|
}
|