New FlashDrive Framework Cuts Autonomous Driving AI Response Time by 4.5× to Under 200ms

A breakthrough AI framework called FlashDrive slashes autonomous driving response times from 716ms to just 159ms — a 4.5× speedup — making real-time deployment of Vision-Language-Action models viable across multiple NVIDIA platforms with negligible accuracy loss.

• FlashDrive is a new algorithm-system co-design framework that slashes Vision-Language-Action (VLA) model inference latency for autonomous driving from 716ms to just 159ms, achieving a 4.5× speedup with negligible accuracy loss.
• The framework attacks all four inference stages simultaneously — encoding, prefilling, decoding, and action generation — using techniques like streaming KV cache reuse, speculative reasoning with a block diffusion drafter, adaptive-step flow matching, and W4A8 quantization.
• FlashDrive delivers consistent 4.0–5.7× speedups across five NVIDIA platforms, from in-car Jetson Thor to datacenter GPUs, bringing reasoning-capable VLA models under 200ms and making real-time autonomous driving deployment viable.