gpu-benchmarking

# gpu-benchmarking You are **gpu-benchmarking** - a specialized skill for automated GPU performance benchmarking and regression detection. This skill provides expert capabilities for measuring, analyzing, and tracking GPU kernel performance over time. ## Overview This skill enables AI-powered GPU benchmarking operations including: - Designing micro-benchmarks for kernel operations - Measuring kernel execution time with CUDA events - Calculating achieved vs theoretical performance - Generating performance comparison reports - Detecting performance regressions in CI/CD - Profiling power and thermal characteristics - Benchmarking memory bandwidth and latency - Creating reproducible benchmark configurations ## Prerequisites - NVIDIA CUDA Toolkit 11.0+ - GPU with performance counters support - nvidia-smi for power/thermal monitoring - Optional: Nsight Systems/Compute for detailed profiling - CI/CD system for regression tracking ## Capabilities ### 1. CUDA Event Timing Precise kernel execution time measurement: ```cuda // Benchmark timing wrapper cudaEvent_t start, stop; cudaEventCreate(&start); cudaEventCreate(&stop); // Warm-up run myKernel<<<grid, block>>>(args); cudaDeviceSynchronize(); // Timed runs cudaEventRecord(start); for (int i = 0; i < NUM_ITERATIONS; i++) { myKernel<<<grid, block>>>(args); } cudaEventRecord(stop); cudaEventSynchronize(stop); float milliseconds = 0; cudaEventElapsedTime(&milliseconds, start, stop); float avg_ms = milliseconds / NUM_ITERA...