Infrastructure-Wide Profiling of NVIDIA CUDA

Learn how to implement infrastructure-wide profiling for NVIDIA CUDA workloads through this 21-minute conference talk from Ubuntu Summit 25.10. Discover the limitations of existing GPU profiling tools like NVIDIA NSight for production environments and explore a groundbreaking solution that enables continuous monitoring of CUDA kernel executions. Understand how the new CUDA kernel execution profiling feature in the Parca open source project traces all CUDA kernel executions, records total execution times, and captures function call-stacks that trigger kernel calls. Explore practical applications for optimizing GPU workloads and see live demonstrations of the profiling system in action. Gain insights from Frederic Branczyk, founder of Polar Signals and former Senior Principal Engineer at Red Hat, who brings extensive experience as a Prometheus and Thanos maintainer and former tech lead for Kubernetes SIG instrumentation. Master the techniques needed to make GPU workload optimization as straightforward as traditional CPU and memory profiling, enabling always-on observability for CUDA applications in production environments.