NVIDIA documentation is sparse at best when it comes to the details of scheduling work on NVIDIA GPUs. Unfortunately, NVIDIA has positioned themselves as an industry leader in GPUs for automotive applications. In order to be used in certifiable safety-critical applications, the scheduling rules that govern how work is ordered on NVIDIA GPUs must be made clear. In this project, we discerned these rules through a series of benchmarking experiments, provided a set of best practices for using NVIDIA GPUs in such safety-critical applications, and developed a software tool to validate whether those best practices are being followed.

Papers:

• GPU Scheduling Rules: Initial Investigation

• GPU Scheduling Rules: Full Investigation

• Pitfalls of using NVIDIA GPUs in Real-Time Applications

• CUPiD^RT: Detecting Improper GPU Usage in Real-Time Applications