Why Deep Learning Needs Assembler Hackers


I doubt you are an assembler hacker (and I'm not linking to this article because I think you should become one). But the author makes a really fascinating observation: there are elements of deep learning algorithms that are not well-optimized by modern compilers. Because of the massive datasets required by these algorithms, the performance implications of even minor optimizations become very significant.

This feels very much to me like when it was still challenging to accept credit card payments online, circa 1998—a great reminder that we are still at the very beginning of this particular era.


Want to receive more content like this in your inbox?