h3-blackwell-runtime/benchmarks/b200-vs-rtxpro6000-server-sdpa-1344x768-seed440420.json

108 lines
2.9 KiB
JSON
Raw Normal View History

{
"device": "NVIDIA B200",
"comparison_device": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
"attention": "sdpa",
"mode": "tensor",
"world_size": 1,
"steps": 12,
"seed": 440420,
"torch": "2.9.1+cu130",
"hourly_usd": {
"b200": 6.79,
"rtx_pro_6000": 2.09
},
"b200_host": {
"data_center": "US-NC-2",
"driver": "580.105.08",
"memory_mib": 183359,
"power_limit_w": 1000
},
"results": [
{
"resolution": [
1344,
768
],
"frames": 124,
"video_seconds_at_24fps": 5.166666666666667,
"b200": {
"model_load_seconds": 11.831760542932898,
"conditioning_seconds": 45.36778333503753,
"sampling_seconds": 79.72125827614218,
"sampling_cost_usd": 0.15036315102639039,
"sampling_peak_allocated_bytes": 17246854656,
"checksums": [
-83553.09375,
-397.89813232421875
],
"console_step_seconds_rounded": [
6.4,
6.1,
6.1,
6.1,
6.1,
6.1,
6.1,
6.1,
6.1,
6.1,
6.1,
6.1
]
},
"rtx_pro_6000": {
"sampling_seconds": 97.94385590963066,
"sampling_cost_usd": 0.05686184968086891
},
"b200_speedup": 1.2285789013812125,
"b200_sampling_cost_ratio": 2.6443591242685138
},
{
"resolution": [
1344,
768
],
"frames": 243,
"video_seconds_at_24fps": 10.125,
"b200": {
"model_load_seconds": 10.712746233213693,
"conditioning_seconds": 38.57332478091121,
"sampling_seconds": 257.8619639207609,
"sampling_cost_usd": 0.48635631528387957,
"sampling_peak_allocated_bytes": 22955644416,
"checksums": [
41606.85546875,
629.4112548828125
],
"console_step_seconds_rounded": [
21.4,
21.1,
21.1,
21.1,
21.0,
21.0,
21.0,
21.0,
21.0,
21.0,
21.0,
21.0
]
},
"rtx_pro_6000": {
"sampling_seconds": 311.01632468774915,
"sampling_cost_usd": 0.18056225516594326
},
"b200_speedup": 1.2061349411862938,
"b200_sampling_cost_ratio": 2.6935657999890426
}
],
"cost_scope": "Sampling wall time only. Model/Qwen loading, text conditioning, VAE decode, media encoding, muxing, and pod startup are excluded.",
"notes": [
"The 124-frame run preceded the 243-frame run on the same host.",
"Both successful runs followed one pre-sampling dependency failure, so host and network-volume file caches were warm.",
"Console step durations are rounded to one decimal place; sampling_seconds is the authoritative synchronized benchmark timer.",
"The RTX PRO 6000 measurements are copied from rtxpro6000-server-1v2-sdpa-scaling-seed440420.json."
]
}