Improving Simplifiable SUSP Simplify Algorithms In CUDA

This presentation introduces a framework to improve the efficiency of matrix multiplication using Uniquely Solvable Puzzles (USP) and a simpler subset called SUSP. First, we explain the basic idea of matrix multiplication and the theory behind USP and SUSP. Then we use simulations to show that SUSP can reduce computational overhead and make the process more efficient. Our work includes a method to detect large SUSP structures and simplify them, together with a search algorithm that selects the best SUSP candidates. To address the limitations of CPU computation, we also design a GPU-accelerated approach. Our analysis shows that this method greatly reduces processing time and scales better for large problems. Finally, we compare CPU and GPU implementations, analyze their performance and bottlenecks, and discuss future directions such as further algorithm improvements, hybrid CPU-GPU systems, and extending the framework to more complex matrix operations and applying machine learning.