How do WorldRoamBench's single-H20 throughput measurements for LingBot-World, HY-World 1.5, Matrix-Game 2.0/3.0 and Yume 1.5 compare with the authors' stated FPS, and what can and cannot be concluded?
WorldRoamBench v5 (AMAP CV Lab, Alibaba Group; not the authors of the open models) timed 10 open models on one NVIDIA H20, 20 runs, as chunk frames over wall-clock chunk time. Measured: Matrix-Game 2.0 6.26 Hz (640x352, 12-frame chunks), Matrix-Game 3.0 5.34 Hz (1280x704, 40), Yume 1.5 6.49 Hz (1280x704, 32), LingBot-World 1.79 Hz (832x480, 12), LingBot-World 2.0 1.75, HY-World 1.5 1.30 Hz (832x480, 16). Against the authors' figures (LingBot 16 fps on one GPU node; Matrix-Game 2.0 25 FPS on one H100 per C3-B062-W13-05) these are lower, but the hardware, GPU count and settings differ, so none of the gaps shows an author figure to be wrong.
WorldRoamBench v5 (AMAP CV Lab, Alibaba Group; not the authors of the open models) timed 10 open models on one NVIDIA H20, 20 runs, as chunk frames over wall-clock chunk time. Measured: Matrix-Game 2.0 6.26 Hz (640x352, 12-frame chunks), Matrix-Game 3.0 5.34 Hz (1280x704, 40), Yume 1.5 6.49 Hz (1280x704, 32), LingBot-World 1.79 Hz (832x480, 12), LingBot-World 2.0 1.75, HY-World 1.5 1.30 Hz (832x480, 16). Against the authors' figures (LingBot 16 fps on one GPU node; Matrix-Game 2.0 25 FPS on one H100 per C3-B062-W13-05) these are lower, but the hardware, GPU count and settings differ, so none of the gaps shows an author figure to be wrong.
Curator qualifications: Third-party throughput on one H20 for open models; the second figure is MPPS (million pixels per second), verified by recomputation (640x352x6.26/1e6 = 1.41). Throughput, not key-press-to-frame latency. Clarification for LEAD-QA: the second number in each row is MPPS = width x height x Hz / 10^6.
Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.
Structured fields
{
"citations": [
{
"url": "https://arxiv.org/abs/2606.31672v5",
"note": "WorldRoamBench v5 efficiency table and definitions"
},
{
"url": "https://arxiv.org/abs/2601.20540v1",
"note": "WorldRoamBench v5 efficiency table and definitions"
},
{
"url": "https://arxiv.org/abs/2604.21686v2",
"note": "WorldRoamBench v5 efficiency table and definitions"
},
{
"url": "https://github.com/Tencent-Hunyuan/Hunyuan-GameCraft-1.0/issues",
"note": "WorldRoamBench v5 efficiency table and definitions"
}
]
}Evidence claims
- [sourced_fact] WorldRoamBench v5 (AMAP CV Lab, Alibaba Group; not the authors of the open models) timed 10 open models on one NVIDIA H20, 20 runs, as chunk frames over wall-clock chunk time. Measured: Matrix-Game 2.0 6.26 Hz (640x352, 12-frame chunks), Matrix-Game 3.0 5.34 Hz (1280x704, 40), Yume 1.5 6.49 Hz (1280x704, 32), LingBot-World 1.79 Hz (832x480, 12), LingBot-World 2.0 1.75, HY-World 1.5 1.30 Hz (832x480, 16). Against the authors' figures (LingBot 16 fps on one GPU node; Matrix-Game 2.0 25 FPS on one H100 per C3-B062-W13-05) these are lower, but the hardware, GPU count and settings differ, so none of the gaps shows an author figure to be wrong. · locator: WorldRoamBench v5 efficiency table and definitions source
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | unknown | unknown | Applies to this original summary only, not to the cited sources. |