TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

chips infrastructure

Four Bottlenecks Fell to One Driving Loop

FlashDrive combined reuse, speculative drafting, adaptive denoising and systems work to cut reported VLA latency from 717 milliseconds to 151.

Published Updated Story ID: mp-2026-08-15-002
Read the complete editionStory JSON

Summary

FlashDrive combined reuse, speculative drafting, adaptive denoising and systems work to cut reported VLA latency from 717 milliseconds to 151.

Vision-language-action driving models have to turn video and language into motion quickly enough for control, but the FlashDrive authors argue that no single optimization fixes the whole delay. Their design reuses visual and key-value context across overlapping frames, drafts low-entropy reasoning tokens non-autoregressively, caches the flatter middle of a flow-matching trajectory, and adds CUDA Graph compilation and kernel fusion.

Applied to a quantized 10-billion-parameter driving model on one GPU, the paper reports a 4.7-fold latency reduction, from 717 milliseconds to 151, raising the loop from 1.4 to 6.6 hertz. The reported trajectory error changed by 0.08 meter on one metric, while single-path error and simulated collision and off-road rates improved. Those safety figures come from closed-loop simulation, not a public-road deployment, and the result is specific to the tested model, hardware and optimization stack.

Why it matters

FlashDrive combined reuse, speculative drafting, adaptive denoising and systems work to cut reported VLA latency from 717 milliseconds to 151.

Limits and context

  • Those safety figures come from closed-loop simulation, not a public-road deployment, and the result is specific to the tested model, hardware and optimization stack.

Key claims

  1. FlashDrive combined reuse, speculative drafting, adaptive denoising and systems work to cut reported VLA latency from 717 milliseconds to 151.

    Qualification: Those safety figures come from closed-loop simulation, not a public-road deployment, and the result is specific to the tested model, hardware and optimization stack.

    Evidence: source-2026-08-15-002

Sources

  1. arXiv preprint 2608.12932arXiv · primary research

Corrections

No corrections have been recorded for this story.