Triton (programming language)

2021: Announcement of the programming language for creating neural networks

At the end of July 2021, OpenAI introduced the Triton open source programming language, which, according to the company, will allow developers to easily create high-speed machine learning algorithms.

Triton was first introduced in 2019 at the International Seminar on Machine Learning and Programming Languages. Soon after, the creator of the programming language, Philip Tille, moved to work in OpenAI.

Triton developers claim that the programming language will achieve maximum hardware performance without much effort. Triton is designed to write high-performance code executed on a graphics processor by developers without experience with CUDA.

OpenAI released the Triton programming language for creating neural networks

The Triton programming language is compatible with Python, the developers managed to perform a task using 25 lines, the implementation of which took much more resources. The solution presented OpenAI also simplifies the creation of specialized cores, which can be much faster than analogues in general-purpose libraries. The compiler automatically optimizes the code by converting it to run on the latest Nvidia graphics processors.

Triton allows you to achieve maximum hardware performance with relatively little effort, "Tille explained. - For example, it can be used to multiply FP16 matrices, which corresponds to cuBLAS performance - which many GPU programmers cannot do in less than 25 lines of code.

Tille also noted that OpenAI researchers have already used Triton to create nuclei that are twice as effective as equivalent Torch implementations. Nevertheless, Triton does not automate the entire workflow, as OpenAI sought to provide developers with the first flexibility to manually customize the process for their projects as needed.^[1]

Notes