hotspot identification

**Hotspot identification** is the **process of locating the small set of operations responsible for most runtime cost** - it applies Pareto-style focus so optimization effort targets the highest-return bottlenecks first. **What Is Hotspot identification?** - **Definition**: Ranking operators or kernels by inclusive and self time contribution to overall step runtime. - **Common Hotspots**: Large GEMM kernels, attention ops, data transforms, and synchronization-heavy collectives. - **Measurement Inputs**: Profiler operator tables, kernel traces, memory counters, and communication metrics. - **Outcome**: Short prioritized list of components for targeted optimization or replacement. **Why Hotspot identification Matters** - **Efficiency**: Most runtime gains come from improving a few dominant hotspots rather than broad micro-tuning. - **Engineering Focus**: Prevents teams from spending time on low-impact parts of the stack. - **Predictable Gains**: Hotspot-driven work usually yields measurable throughput improvements quickly. - **Resource Allocation**: Supports better prioritization of compiler, kernel, and model-engineering effort. - **Continuous Improvement**: Repeated hotspot analysis tracks how bottlenecks shift after each optimization round. **How It Is Used in Practice** - **Top-N Extraction**: Generate ranked runtime table and isolate the smallest set covering most total time. - **Constraint Classification**: Determine whether each hotspot is compute, memory, communication, or launch overhead bound. - **Focused Optimization**: Apply specialized fixes to top hotspots and validate runtime share reduction afterward. Hotspot identification is **the practical starting point for performance optimization campaigns** - optimizing the right few kernels delivers far more value than tuning everything equally.