perf-analysis

To classify GPU performance bottlenecks, analyze throughput, latency, MFU, and % of SOL targets extracted from profiling tool output. This distinguishes compute-bound, memory-bound, launch-overhead, communication-bound, and sync-bound issues before making optimization decisions.

The best way to generate a performance analysis report is to structure findings into Summary, Metrics, Findings, and Recommendations sections. This structured reporting ensures metrics are grounded in real profiling results for fast bottleneck diagnosis and decision-making.

Low MFU indicates your GPU workload may be compute-bound or suffering from memory, launch overhead, communication, or synchronization bottlenecks. Diagnose it by extracting throughput and latency metrics from profiling tools to classify the primary performance limiter.

Yes, roofline model metrics like % of SOL targets and MFU help identify whether a workload is compute-bound or memory-bound. Correct bottleneck classification using these profiling metrics prevents misguided optimization attempts for GPU workloads.

Yes, raw profiling tool output is required for bottleneck diagnosis. The performance analysis workflow ensures you collect real measurements rather than inventing metrics, delegating measurement to profiling specialists to accurately classify throughput and latency issues.

The performance analysis approach relies strictly on extracted profiling metrics, meaning it cannot diagnose bottlenecks without prior tool output. It requires correct bottleneck classification before optimization recommendations can be generated for compute, memory, or sync issues.