gpu-kernel-bottleneck-analysis

To diagnose if a CUDA kernel is compute-bound or memory-bound, calculate its TFLOPS and bandwidth utilization against a hardware-specific Roofline model. This skill extracts concrete metrics from NVIDIA NCU profiling data to pinpoint the specific bottleneck.

GPU Roofline analysis is a diagnostic framework that measures compute and memory utilization against theoretical hardware peaks. By parsing profiling summaries, it identifies specific ISA-level stalls, bank conflicts, or low occupancy preventing higher throughput.

Yes, you can profile AMD ROCprof data to identify kernel underperformance. The analysis supports diagnostic workflows for both compute-bound and memory-bound kernels by parsing ROCprofiling summary files to extract actionable optimization evidence.

To find ISA-level bottlenecks, execute the provided Python benchmarking tools to generate profiling data, then parse the summary files. This evidence-based approach moves beyond guesswork to pinpoint exact stalls or low occupancy issues.

Yes, you need to run Python benchmarking tools to generate the profiling summary files required for analysis. The skill parses these extracted metrics to determine if the kernel is memory-bound or compute-bound against the hardware Roofline.