Machine Learning Surgery’s New Instrument: Triton

#

One of the most amazing things about being a Machine Learning Surgeon is that you never get bored at work. The reason? There’s always novelty, always new things to learn and discover, mainly because your colleagues are always busy creating new exciting things.

A fairly new instrument is called Triton. Triton, developed and maintained by OpenAI, is by definition a language and a compiler for parallel programming. It was announced circa the end of 2021 and since then it grew a large community of users, thanks to its simplicity and power.

Its easy of use is mainly due to the fact that Triton allows developers to write code at a higher level of abstraction, while still achieving performance comparable to hand-optimized CUDA code. It does this by automatically handling many of the complex optimizations that GPU programmers typically need to manage manually such as memory access patterns and thread blocking.

This is a game changer, since many developers never approached GPU programming because of CUDA’s steep learning curve and