How an autonomous trucking company reduced perception deployment costs by 30% with yasp

Autonomous trucking company eliminates per-target kernel engineering and speeds up binaries for AGX Orin by 2x

Published on

Jun 2026

Summary

The company deploys real-time perception models on NVIDIA DRIVE hardware inside autonomous trucks. yasp reduced their monthly compute spend by approximately 30%, eliminated weeks of per-target kernel engineering on each hardware transition, and compressed model deployment timelines from months to weeks.

Impact

2-4x

Speedup

Months to Days

Time to market

30%

Cost reduction across perception workloads

"We spent years building AI infrastructure and kept hitting the same wall: a model that works everywhere in theory but only runs where the vendor lets it. yasp is our answer to that. You choose the hardware that's right for the job, our agents handle the optimization and deployment automatically, and the model runs because nothing stands in the way. You choose the hardware. yasp handles the rest."

Stefan Krassin

VP Engineering

,

yasp

Challenge

The autonomous vehicle development team was managing significant compute costs for its perception pipeline. Model training ran in the cloud, while inference ran on embedded NVIDIA hardware at the edge.

Several custom operations within the sensor-fusion pipeline required hand-written CUDA kernels optimized for specific hardware. Each hardware-generation change required kernel updates, regression testing, and revalidation, creating substantial deployment and maintenance overhead.

Solution

The team began with a high-maintenance sensor-fusion module and compiled it through yasp for an NVIDIA embedded computing platform.

The custom sensor-fusion operations compiled natively, and the resulting binary delivered performance comparable to or better than the existing optimized deployment path. The team then expanded the approach to additional perception and prediction models, with similarly consistent results.

Impact

yasp sped up the binaries for NVIDIA DRIVE AGX Orin by 2x, creating a 30% reduction across the perception pipeline. The migration to AGX Orin, which had been budgeted as a quarter-long project, took days.

Model updates that previously stalled in deployment engineering now ship in the same sprint as retraining. The team is preparing for NVIDIA DRIVE AGX Thor. A hardware transition that would previously have triggered months of kernel migration is now a compile target.