Efficient Matrix Multiplication on Homogenous GPUs Using Reinforcement Learning for Scientific Computing

Matrix multiplication serves as a critical operation in neural networks and scientific computing, where algorithmic improvements can significantly impact execution speed. Existing optimized algorithms still leave room for performance enhancement. This paper proposes a reinforcement learning (RL) framework to dynamically select matrix multiplication methods based on matrix attributes and hardware conditions. A dataset of matrices with varying dimensions and densities is used to train an RL agent operating on homogeneous GPUs, optimizing algorithm selection to enhance parallel processing efficiency. Real-time performance metrics, including computational complexity, runtime, temperature, and voltage, guide decision-making. The proposed approach demonstrates how RL can substantially reduce computation time and improve resource utilization for large-scale matrix operations. Experimental evaluation validates the efficiency of RL-based optimization in scientific computing tasks, emphasizing practical improvements in computational speed and hardware-aware resource management.