Event
Seven-Stage Latency Benchmark Study for VLA Edge Inference
1 reports · 1 sources
Borui Xu and other authors released VLA-HWBench, which performs a seven-stage inference decomposition of 18 open-source VLA models on NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395, finding that action generation strategy determines latency more than model scale, with parameter count explaining at most 15% of latency variance. The benchmark states that models generating long action chunks with a small number of flow-matching steps can keep up with the control rate, whereas token-by-token decoding cannot. The authors also note that the results cover only these two hardware platforms, and that they exist only on Jetson.
Report timeline
- VLA-HWBench: Where inference time goes for VLA models on edge hardware
VLA-HWBench breaks down inference into seven stages for 18 open-source VLA models on NVIDIA Jetson AGX Thor and AMD Ryzen AI Max+ PRO 395, finding that action-generation strategy determines latency more than model size does, with parameter count explaining at most 15% of latency variance.
arXiv cs.RO ·