Training Throughput Measurement on ManiSkill
436.32Training ThroughputTrain Async
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Train AsyncVLA Backbone=pi0, Number of GPUs=32 GPUs, GPU Allocation Ratio=1:12026.02 | 436.32 | — | |
| ColocatedVLA Backbone=pi0, Number of GPUs=32 GPUs2026.02 | 370.26 | — | |
| StreamerVLA Backbone=pi0, Number of GPUs=32 GPUs, GPU Allocation Ratio=1:12026.02 | 280.07 | 17.84 | |
| Rollout AsyncVLA Backbone=pi0, Number of GPUs=32 GPUs, GPU Allocation Ratio=1:12026.02 | 275.36 | — | |
| DisaggregatedVLA Backbone=pi0, Number of GPUs=32 GPUs, GPU Allocation Ratio=1:12026.02 | 257.21 | — | |
| Train AsyncVLA Backbone=pi0, Number of GPUs=16 GPUs, GPU Allocation Ratio=1:12026.02 | 244.61 | — | |
| ColocatedVLA Backbone=pi0, Number of GPUs=16 GPUs2026.02 | 232.23 | — | |
| DisaggregatedVLA Backbone=pi0, Number of GPUs=16 GPUs, GPU Allocation Ratio=1:12026.02 | 150.59 | — | |
| StreamerVLA Backbone=pi0, Number of GPUs=16 GPUs, GPU Allocation Ratio=1:12026.02 | 141.49 | 5.33 | |
| Rollout AsyncVLA Backbone=pi0, Number of GPUs=16 GPUs, GPU Allocation Ratio=1:12026.02 | 139.32 | — | |
| ColocatedVLA Backbone=pi0, Number of GPUs=8 GPUs2026.02 | 132.56 | — | |
| Train AsyncVLA Backbone=pi0, Number of GPUs=8 GPUs, GPU Allocation Ratio=1:12026.02 | 126.65 | — | |
| StreamerVLA Backbone=pi0, Number of GPUs=8 GPUs, GPU Allocation Ratio=1:12026.02 | 71.61 | -4.46 | |
| Rollout AsyncVLA Backbone=pi0, Number of GPUs=8 GPUs, GPU Allocation Ratio=1:12026.02 | 69.07 | — | |
| DisaggregatedVLA Backbone=pi0, Number of GPUs=8 GPUs, GPU Allocation Ratio=1:12026.02 | 60.64 | — |