It was unique that matrix multiplication can be implemented in various ways, and I didn't know that changing the loop structure in CPU implementation would make it faster. It feels like I learned practical coding, not the level that usually appears in language books.