perf-nsight-compute-analysis

To identify CUDA kernel bottlenecks from an Nsight Compute report, the Skill analyzes .ncu-rep files using SOL% classification to determine if kernels are compute-bound, memory-bound, or latency-bound.

SOL% based bottleneck classification in GPU profiling measures the speed of light to identify whether a CUDA kernel is compute-bound, memory-bound, or latency-bound, guiding subsequent optimization priorities.

Roofline analysis guides CUDA kernel optimization by mapping workload data against hardware limits in the .ncu-rep report, pinpointing whether memory bandwidth or compute throughput restricts performance.

Yes, you need .ncu-rep files generated by Nsight Compute to analyze kernel performance, as these reports contain the required SpeedOfLight and workload analysis sections for consistent results.

For memory workload analysis, you should deep dive into the MemoryWorkloadAnalysis and ComputeWorkloadAnalysis sections of the .ncu-rep report to diagnose memory hierarchy bottlenecks.

Yes, this profiling workflow supports different GPU architectures by providing reproducible guidance for analyzing Nsight Compute profiling outputs across various versions and hardware architectures.