2B Parameter VLA Integrated into In-Vehicle Systems: NVIDIA Thor Platform Test Shows Minimum Latency of 160ms
NVIDIA has released the Alpamayo 2 Edge model adapted for the DRIVE AGX Thor in-vehicle platform. This 2 billion parameter-level inference VLA can directly perform driving trajectory planning and decision interpretation on the vehicle end, with an end-to-end inference latency of 160 to 190 milliseconds from pixels to trajectory.
On November 22, Asif Razzaq, a writer for tech media MarkTechPost, disclosed NVIDIA's latest progress in in-vehicle large models.
Named Alpamayo 2 Edge, this model is a 2B (2 billion parameter) level inference Vision-Language-Action (VLA) model, specifically adapted to NVIDIA's own DRIVE AGX Thor in-vehicle computing platform. It has two core functions: first, it directly outputs driving trajectory planning based on pixel data collected by in-vehicle cameras; second, it simultaneously provides logical explanations for its decisions. The measured end-to-end latency, which refers to the total time from pixel input to trajectory output, falls between 160 and 190 milliseconds.
For readers without technical backgrounds, let’s first add two basic pieces of context. VLA is a sub-direction of multimodal large models, which has the capabilities of image understanding, language processing, and action output simultaneously. In the autonomous driving scenario, the model can not only understand pedestrians and vehicles on the road, recognize traffic signs, but also understand language instructions, and directly output corresponding driving operation paths. Previously, most VLAs with complete decision-making capabilities had a large parameter scale and could only be deployed in cloud data centers. Vehicles need to upload the collected road surface data, wait for the cloud calculation to complete, and then transmit the decision results back. The entire process relies on network signals, and it is difficult to reduce the latency to meet the safety requirements of in-vehicle systems.
DRIVE AGX Thor is an automotive-grade computing platform launched by NVIDIA for high-level autonomous driving, designed to carry the inference tasks of large models on the vehicle end. This 2B parameter VLA achieved a latency of 160 milliseconds on the Thor platform, which happens to fall within the safety range for real-time in-vehicle decision-making. In high-speed driving scenarios, a delay difference of tens of milliseconds corresponds to a braking distance of several meters. The combination of local vehicle-side operation and interpretable decision-making features directly addresses two common pain points of current autonomous driving solutions: the instability caused by cloud dependence and the black-box problem of opaque decision-making processes.
Currently, NVIDIA has not yet disclosed the timeline for mass production and vehicle installation of this model.
发布时间: 2026-11-23 02:37