GTC Silicon Valley-2019 ID:S9652:Achieving Deterministic Execution Times in CUDA Applications
Ashok Kelur(NVIDIA),Aayush Rajoria(NVIDIA)
CUDA has been an industry standard for high-performance computing applications to use GPU parallelism for general-purpose computing. Achieving high compute throughput has always been an important goal. But CUDA is increasingly used in autonomous vehicles and robotics, where deterministic execution time is important. We'll discuss some application and system design considerations to help CUDA developers achieve deterministic execution times. We will also talk about tricks to avoid bubbles in the GPU pipeline and improve GPU utilization, general programming practices, and application design for deterministic execution times.