gpu-perf-prophet / data /models /feature_metadata.json
vxa8502's picture
Close power-efficiency Pareto gap, add training-data disclosure, harden data lineage, cut Docker image 84%
3c35b12
Raw
History Blame Contribute Delete
7.6 kB
{
"feature_cols": [
"gpu_hbm_bandwidth_tbps",
"gpu_vram_gb",
"peak_tflops_selected",
"compute_ceiling_tok_per_sec",
"bandwidth_ceiling_tok_per_sec",
"model_total_params_b",
"model_compute_params_b",
"model_size_gb",
"model_to_vram_ratio",
"bytes_per_param",
"nvidia_arch_gen",
"amd_arch_gen",
"vendor_is_amd",
"scenario_offline",
"is_base_tier",
"fw_tensorrt",
"fw_vllm",
"fw_rocm_other",
"is_cdna4",
"mlperf_round_num"
],
"target": "efficiency_ratio",
"model_version": "v1",
"prod_params": {
"n_estimators": 250,
"max_depth": 5,
"learning_rate": 0.03,
"subsample": 0.8,
"colsample_bytree": 0.8,
"min_child_weight": 5,
"tree_method": "hist",
"random_state": 42
},
"n_training_rows": 1136,
"corpus_sha256": "398521848da48ab557ab611104e994fa3261f23f33ae5f7ee26c5cca3093139c",
"corpus_path": "data/processed/mlperf_features.parquet",
"manifest_verified": true,
"manifest_sources_present": 5,
"manifest_sources_total": 5,
"trained_gpu_ids": [
"b200_sxm",
"b300_sxm",
"gb200",
"gb300",
"gh200_144gb",
"gh200_96gb",
"h100_nvl",
"h100_pcie",
"h100_sxm",
"h200_nvl",
"h200_sxm",
"l40s",
"mi300x",
"mi325x",
"mi350x",
"mi355x"
],
"trained_gpu_row_counts": {
"h200_sxm": 283,
"h100_sxm": 178,
"b200_sxm": 138,
"mi325x": 82,
"mi300x": 80,
"h200_nvl": 52,
"gh200_144gb": 50,
"mi355x": 50,
"h100_nvl": 46,
"l40s": 45,
"b300_sxm": 29,
"gb200": 28,
"gh200_96gb": 26,
"h100_pcie": 24,
"mi350x": 14,
"gb300": 11
},
"validation_metrics": {
"generated_at": "2026-07-12T13:07:49.619349+00:00",
"source_notebook": "notebooks/03_model_training.ipynb",
"protocol": "LOGO-CV (Leave-One-GPU-Out); primary = in-scope folds (n>=10); official-only scoring where a fold mixes official MLPerf submissions with self-run calibration rows",
"primary": {
"n_folds": 5,
"mape": 0.23380974767055468,
"smape": 0.21164078416451165,
"median_ape": 0.18509508530648328,
"spearman_rho": 0.7888808863450961,
"worst_fold_rho": {
"gpu": "mi355x",
"rho": 0.5121800160905657
},
"worst_fold_mape": {
"gpu": "mi300x",
"mape": 0.4496798476393815
}
},
"vendor": {
"nvidia": {
"mape": 0.21373571232654986,
"smape": 0.23002982188975515,
"median_ape": 0.19212369651078004,
"rho": 0.8939945816961098,
"n_folds": 2
},
"amd": {
"mape": 0.24719243789989123,
"smape": 0.199381425681016,
"median_ape": 0.18040934450361876,
"rho": 0.7188050894444205,
"n_folds": 3
}
},
"per_fold": [
{
"gpu": "b200_sxm",
"vendor": "nvidia",
"n": 138,
"in_scope": false,
"in_agg": true,
"mape": 0.3374755307879345,
"smape": 0.29186819077474774,
"median_ape": 0.23694338395783016,
"spearman_rho": 0.8306123211824746
},
{
"gpu": "b300_sxm",
"vendor": "nvidia",
"n": 29,
"in_scope": false,
"in_agg": true,
"mape": 0.23684619061590245,
"smape": 0.2564624394530722,
"median_ape": 0.27144706467076973,
"spearman_rho": 0.818841138208717
},
{
"gpu": "gb200",
"vendor": "nvidia",
"n": 28,
"in_scope": false,
"in_agg": true,
"mape": 0.3098156351506685,
"smape": 0.2729216244084394,
"median_ape": 0.2036609681628041,
"spearman_rho": 0.7650111899500084
},
{
"gpu": "gb300",
"vendor": "nvidia",
"n": 11,
"in_scope": false,
"in_agg": true,
"mape": 0.17430911204008873,
"smape": 0.17393644610376116,
"median_ape": 0.14098983643432073,
"spearman_rho": 0.8571428571428571
},
{
"gpu": "gh200_144gb",
"vendor": "nvidia",
"n": 50,
"in_scope": false,
"in_agg": true,
"mape": 0.06831530675958356,
"smape": 0.07069476117842656,
"median_ape": 0.06896169387469567,
"spearman_rho": 0.8917469312909448
},
{
"gpu": "gh200_96gb",
"vendor": "nvidia",
"n": 26,
"in_scope": false,
"in_agg": true,
"mape": 0.09661322592006306,
"smape": 0.09941079345148354,
"median_ape": 0.09330221054708361,
"spearman_rho": 0.9177971765602133
},
{
"gpu": "h100_nvl",
"vendor": "nvidia",
"n": 46,
"in_scope": false,
"in_agg": true,
"mape": 0.5456500368887637,
"smape": 0.41980135638955235,
"median_ape": 0.4683624244793203,
"spearman_rho": 0.9145629925449604
},
{
"gpu": "h100_pcie",
"vendor": "nvidia",
"n": 24,
"in_scope": false,
"in_agg": true,
"mape": 0.3647269379714631,
"smape": 0.30970163452636484,
"median_ape": 0.24785922361578588,
"spearman_rho": 0.8454148471615721
},
{
"gpu": "h100_sxm",
"vendor": "nvidia",
"n": 178,
"in_scope": true,
"in_agg": true,
"mape": 0.23675355291273484,
"smape": 0.28349726152128213,
"median_ape": 0.23749937063846838,
"spearman_rho": 0.8961478340151705
},
{
"gpu": "h200_nvl",
"vendor": "nvidia",
"n": 52,
"in_scope": false,
"in_agg": true,
"mape": 0.23680727903776597,
"smape": 0.19338038678086072,
"median_ape": 0.16766815365958276,
"spearman_rho": 0.8654862039709149
},
{
"gpu": "h200_sxm",
"vendor": "nvidia",
"n": 283,
"in_scope": true,
"in_agg": true,
"mape": 0.1907178717403649,
"smape": 0.17656238225822818,
"median_ape": 0.1467480223830917,
"spearman_rho": 0.8918413293770491
},
{
"gpu": "l40s",
"vendor": "nvidia",
"n": 45,
"in_scope": false,
"in_agg": true,
"mape": 0.9042459165863032,
"smape": 0.5572620474803024,
"median_ape": 0.6657441551562934,
"spearman_rho": 0.6975573825353433
},
{
"gpu": "mi300x",
"vendor": "amd",
"n": 80,
"in_scope": true,
"in_agg": true,
"mape": 0.4496798476393815,
"smape": 0.32283498644102115,
"median_ape": 0.33939678840089643,
"spearman_rho": 0.7452002548867019
},
{
"gpu": "mi325x",
"vendor": "amd",
"n": 82,
"in_scope": true,
"in_agg": true,
"mape": 0.1590707712715832,
"smape": 0.1279615122994294,
"median_ape": 0.058077964040084584,
"spearman_rho": 0.8990349973559938
},
{
"gpu": "mi350x",
"vendor": "amd",
"n": 14,
"in_scope": false,
"in_agg": true,
"mape": 0.13799850100450334,
"smape": 0.11752693059503805,
"median_ape": 0.034923866028296384,
"spearman_rho": 0.01857182112864787
},
{
"gpu": "mi355x",
"vendor": "amd",
"n": 50,
"in_scope": true,
"in_agg": true,
"mape": 0.13282669478870904,
"smape": 0.14734777830259735,
"median_ape": 0.14375328106987523,
"spearman_rho": 0.5121800160905657
}
]
}
}