> It’s arguably why deep learning took off: matrix multiplication happens to be really cheap on GPUs Multiplication of large matrices is intrinsically simple t

> It’s arguably why deep learning took off: matrix multiplication happens to be really cheap on GPUs Multiplication of large matrices is intrinsically simple to parallelize, and in addition it has great data locality and a very regular structure. It's just a great operation to perform efficiently on a parallel processor.