perf-nsight-compute-analysis

To diagnose CUDA kernel underperformance, analyze Nsight Compute SOL% metrics to classify bottlenecks as compute-bound, memory-bound, or latency-bound. Use targeted ncu profiling sections like SpeedOfLight, Occupancy, and Warp/Instruction stats to guide root-cause investigation of GPU kernel issues.

Run ncu with targeted section flags like Occupancy and Warp/Instruction stats to investigate latency and underutilization causes. This section-first workflow avoids exhaustive profiling, focusing only on necessary data to interpret warp stalls and instruction issues efficiently.

Yes, you can extract and classify metrics from saved .ncu-rep files using the ncu_report Python API. This approach enables programmatic analysis of Nsight Compute reports to classify SOL% bottlenecks and diagnose kernel performance issues.

Yes, roofline analysis charts compute-memory balance to classify SOL% bottlenecks. Combined with occupancy and stall-related sections, it visually confirms whether your CUDA kernel is compute-bound or memory-bound, guiding deeper latency root-cause investigation.

SOL% analysis is limited to kernel-level performance diagnosis and does not support nsys system profiling or kernel code generation. It requires running ncu with targeted section flags and focuses strictly on SOL% compute-versus-memory classification, occupancy, and warp stall interpretation.