OpenAI has released an open-source, Python-like programming language called Triton. This will allow researchers to write accurate and efficient GPU code for AI workloads. Triton will be able to help in reaching optimum hardware performance with relatively lesser effort as claimed by OpenAI. This is because Triton can produce codes within the same time as an expert would take and that too in lines as few as 25.
Deep neural networks are an important model for AI and it helps in achieving high performances in language processing and computer vision and others. What makes it stronger is its hierarchical structure. CUDA and OpenCL frameworks have increased the performance of programs easier in recent years. Then too GPUs are hard to optimize mainly due to the rapid changes in their architecture.
Even though domain-specific languages and compilers are addressing this problem, the system is less flexible and takes more time than some of the best-handwritten computer kernels. Triton the new open-source will help in overcoming these obstacles so that the focus will remain on the high-level logic of the developer’s code. Philippe Tillet, the original creator of Triton claims that they have already used the language to produce more efficient kernels. He added that they look forward to “make GPU programming more accessible to everyone”.
The idea behind Triton is originally found in a 2019 paper that was submitted to the International Workshop on Machine Learning and Programming Languages. It can produce specialized kernels which are more efficient than the ones in general-purpose libraries as OpenAI claims.