Researchers have found a way to run heavyweight self-driving AI models without maxing out a car's GPU.
Vision-language-action (VLA) models are the latest push in autonomous driving: single neural networks that fold perception, reasoning, and planning into one system instead of stitching together separate modules. That consolidation frees up CPU capacity that used to handle those planning functions, but it piles even more work onto the GPU, which already runs the visual encoder and the model's core reasoning. A new framework splits VLA models at the block layer, keeping the visual encoder and the front portion of the language model on the GPU while shifting the back half to the CPU through an asynchronous pipeline that spans multiple frames. Tested on two driving VLA models, Orion and MindDrive, on the Bench2Drive benchmark, the approach cut average latency by 21.7% and 30.9% respectively, and dropped Orion's estimated peak GPU memory use from 45GB to 29GB.
The real test came when Orion ran alongside a full autonomous-driving stack, Autoware.Universe, on actual vehicle hardware. The unmodified version couldn't run at all because it blew past the onboard GPU memory budget, while the hybrid version ran successfully. That's the gap between a benchmark chart and a shippable product: plenty of VLA driving models look impressive on paper but won't fit the compute budget of an actual car.
Squeezing more performance out of existing hardware is genuinely useful, but it also underscores a problem automakers haven't solved: they still haven't shipped enough GPU to run these models comfortably in the first place.