Real-time VLA on MacBook
Jialin Lu · luxxxlucy.github.io
April 2026 · in progress
Target
Run π0 (and π0.5) on Mac in realtime rates.
Dexmal showed that we can run π0 with 30 Hz on a RTX 4090 —
see Running VLAs at Real-time Speed v1 and v2.
Right now my numbers are based on my quite lame M4 MacBook Air.
RTX 4090 is a much more powerful card.
Current Status
Worst case: If M4 cannot get there, we will fallback to the dirty road: move to a higher-spec M5 Pro/Max chip. The addition of neural accelerator and much wider memory bandwidth shall get us there.
| State | Rate | |
|---|---|---|
| Baseline (Vanilla PyTorch) | ≈ 0.8 Hz | |
| Current | ≈ 2.4 Hz | ← γγΎγγ |
| Near-term target | 5 Hz | |
| Target | 30 Hz |