VLA Benchmark
| model | BPU core num | qtype | all (ms) | pre-process (ms) | vision (ms) | language (ms) | action (ms) | post-process (ms) | memory (GB) |
|---|---|---|---|---|---|---|---|---|---|
| Pi0 | vision 1 language 4 action 4 | vision w8 language w8 action w8 | 98 | 1.5 | 18.5 | 39 | 38.5 | 0.05 | 4.4 |
Annotation:
-
all: The total latency of a single on-device run, including five stages: preprocess, vision, language, action, and postprocess.
-
vision: The latency measured for the 3-core BPU performing parallel inference on three 224×224 images.
On This Page