Spaces:
Running
Running
| { | |
| "schema_version": 1, | |
| "title": "MiniMax-H3 on GB300", | |
| "workload": { | |
| "model": "MiniMax-H3", | |
| "task": "Reference to video + audio", | |
| "width": 1024, | |
| "height": 768, | |
| "frames": 277, | |
| "steps": 50, | |
| "seed": 7, | |
| "prompt": "Published hero prompt; identical across configurations", | |
| "input": "One reference image; identical across configurations" | |
| }, | |
| "environment": { | |
| "hardware": "One GB300 tray · 4 B300 GPUs · Grace aarch64", | |
| "runtime": "SGLang Diffusion", | |
| "runtime_revision": null, | |
| "driver": "595.71.05 (reported in the article)" | |
| }, | |
| "energy_scope": "four_gpu_domains_submit_to_done", | |
| "measurement_notes": [ | |
| "Energy sums four GPU power domains, including idle neighbours, from 500 ms power polling. The capture integrates samples inside the submit-to-done window without filling its edges. Error cannot be reconstructed from the retained TSVs; CPUs, NVSwitches, DPUs, and PSU losses are excluded." | |
| ], | |
| "source": { | |
| "repository": "https://github.com/EzgiTastan/minimax-h3-gb300-ebpf", | |
| "revision": "8734f72a33426f32331b1baf43d0dc7965978a7b", | |
| "article": "https://ezgitastan.systems/blog/minimax-h3-sglang-gb300-ebpf/" | |
| }, | |
| "configurations": [ | |
| { | |
| "id": "one-bf16", "name": "1 GPU · BF16", "gpu_count": 1, | |
| "precision": "BF16", "parallelism": "Single GPU", "source_path": "figdata/rung1.tsv", | |
| "samples": [ | |
| {"run": 0, "infer_s": 336.549, "wall_s": 340.177, "peak_mb": 133270.0, "energy_kj": 694.11, "mean_tray_w": 2039.5}, | |
| {"run": 1, "infer_s": 336.623, "wall_s": 340.151, "peak_mb": 133270.0, "energy_kj": 694.03, "mean_tray_w": 2042.3}, | |
| {"run": 2, "infer_s": 336.589, "wall_s": 340.155, "peak_mb": 133320.0, "energy_kj": 693.96, "mean_tray_w": 2042.1} | |
| ] | |
| }, | |
| { | |
| "id": "two-bf16", "name": "2 GPUs · Ulysses SP2", "gpu_count": 2, | |
| "precision": "BF16", "parallelism": "Ulysses SP2", "source_path": "figdata/rung_sp2.tsv", | |
| "samples": [ | |
| {"run": 0, "infer_s": 177.067, "wall_s": 180.108, "peak_mb": 122020.0, "energy_kj": 559.19, "mean_tray_w": 3100.0}, | |
| {"run": 1, "infer_s": 177.056, "wall_s": 180.102, "peak_mb": 114944.0, "energy_kj": 559.07, "mean_tray_w": 3108.0}, | |
| {"run": 2, "infer_s": 177.291, "wall_s": 180.101, "peak_mb": 114846.0, "energy_kj": 559.22, "mean_tray_w": 3108.8} | |
| ] | |
| }, | |
| { | |
| "id": "four-bf16", "name": "4 GPUs · Ulysses SP4", "gpu_count": 4, | |
| "precision": "BF16", "parallelism": "Ulysses SP4", "source_path": "figdata/rung2.tsv", | |
| "samples": [ | |
| {"run": 0, "infer_s": 90.304, "wall_s": 94.381, "peak_mb": 101330.0, "energy_kj": 486.64, "mean_tray_w": 5153.7}, | |
| {"run": 1, "infer_s": 90.472, "wall_s": 94.046, "peak_mb": 95932.0, "energy_kj": 486.83, "mean_tray_w": 5183.2}, | |
| {"run": 2, "infer_s": 90.397, "wall_s": 94.046, "peak_mb": 95840.0, "energy_kj": 487.15, "mean_tray_w": 5186.6} | |
| ] | |
| }, | |
| { | |
| "id": "four-fp8", "name": "4 GPUs · SP4 + FP8", "gpu_count": 4, | |
| "precision": "FP8 weights", "parallelism": "Ulysses SP4", "source_path": "figdata/rung3.tsv", | |
| "samples": [ | |
| {"run": 0, "infer_s": 87.088, "wall_s": 90.052, "peak_mb": 64742.0, "energy_kj": 464.98, "mean_tray_w": 5170.9}, | |
| {"run": 1, "infer_s": 87.221, "wall_s": 90.049, "peak_mb": 65576.0, "energy_kj": 466.04, "mean_tray_w": 5154.1}, | |
| {"run": 2, "infer_s": 87.191, "wall_s": 90.050, "peak_mb": 64876.0, "energy_kj": 465.69, "mean_tray_w": 5178.7} | |
| ] | |
| } | |
| ] | |
| } | |