VLA Training Throughput (GR00T N1.5) on LIBERO
1,620.39Throughput (samples/s)Rollout Async
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Rollout AsyncNumber of GPUs=322026.03 | 1,620.39 | 43.96 | |
| StreamerNumber of GPUs=322026.03 | 1,592.4 | — | |
| ColocatedNumber of GPUs=322026.03 | 1,125.62 | — | |
| Train AsyncNumber of GPUs=322026.03 | 951.33 | — | |
| Rollout AsyncNumber of GPUs=162026.03 | 816.48 | 19.99 | |
| StreamerNumber of GPUs=162026.03 | 816.48 | 19.99 | |
| Disaggregated (1:1)Number of GPUs=322026.03 | 729.98 | — | |
| ColocatedNumber of GPUs=162026.03 | 680.46 | — | |
| Train AsyncNumber of GPUs=162026.03 | 477.2 | — | |
| StreamerNumber of GPUs=82026.03 | 439.64 | 18.25 | |
| Rollout AsyncNumber of GPUs=82026.03 | 434.2 | — | |
| Disaggregated (1:1)Number of GPUs=162026.03 | 409.6 | — | |
| ColocatedNumber of GPUs=82026.03 | 371.8 | — | |
| Train AsyncNumber of GPUs=82026.03 | 243.57 | — | |
| Disaggregated (1:1)Number of GPUs=82026.03 | 220.81 | — |