Worked on performance optimization for the StreamHPC/rocm-libraries repository, focusing on device grouped GEMM operations. Delivered an asynchronous memory copy feature by refactoring the memory transfer path to use hipMemcpyAsync in place of hipMemcpyWithStream, enabling CPU and GPU operations to overlap and reducing data transfer stalls. This targeted update, implemented in C++ with expertise in CUDA and GPU programming, improved throughput potential for GEMM workloads on AMD GPUs. The work maintained minimal API impact, supporting future tuning and maintainability, and aligned with ROCm’s high-performance computing goals. No critical bugs were addressed, as the primary focus was on performance enhancement.