Amir Zadeh developed a performance benchmarking enhancement for inference kernels in the ROCm/FBGEMM repository. He introduced a warm-up method to stabilize timing and integrated the Kineto profiler, enabling more accurate measurement of kernel execution time and bandwidth. By reducing measurement overhead, his work improved the reliability of benchmarking results and provided actionable data for performance tuning and optimization. Amir utilized C++ and Python, applying skills in GPU computing, profiling, and performance optimization. The depth of his contribution lies in addressing the challenges of precise performance measurement, ultimately supporting more informed optimization decisions for inference workloads in GPU environments.