Intel Technical Director, Greg Laverter, proposed using large language models (LLM) to convert an outdated code written for the NVIDIA CUDA platform to make it compatible with other AI accelerators, such as Gaudi2 or GPU Max from Intel.
During his key report on innovation, Laverter expressed the idea of using LLM and technology, such as Copilot, to teach the Machine Learning model to convert CUDA code into SYCL, a programming tool focused on accelerators from Intel.
One of the problems that Intel, AMD and other accelerators are faced with the advancement of their equipment for AI is the need to reactor the code written for the CUDA execution from NVIDIA before using it on alternative platforms.
Intel has already reached some progress in this direction, actively investing in its cross -platform model of parallel programming OneAPI and a proposal for the output of AI called Openvino. Lavender noted that the OneAPI installation base has increased by 85% since 2021, which indicates a growing interest in the platform.
SYCL is part of the Intel effort to break CUDA in the Ecosystem of the AI software. SYCL, or rather Syclomatic, is a free layer of abstraction, which maintains a parallel C ++ programming language from Intel.
Although SYCL greatly simplifies the CUDA porting process to work on accelerators not related to NVIDIA, additional settings are usually required to achieve maximum performance.
Laverter suggested that this setting process can be automated using the LLM model. “This is not just a good idea, this is an idea that will certainly happen,” said Kurli, vice president of software products and an ecosystem on Intel.
Nevertheless, it is worth noting that there are other ways of writing code, which does not depend on a particular accelerator. For example, the framework of Openai’s Triton or Google’s Jax.
against the background of software implementation media such as SYCL, Intel provides many resources in the form of software, support and accelerators in its Developer Cloud to optimize code for Gaudi2, GPU Max and other products.
This week, Linux Foundation in partnership with a number of companies, including Intel, created a unified acceleration fund (UXL). The purpose of the working group is to develop OneAPI for accelerators from various manufacturers.
According to Laverns, “the industry benefits from open standardized programming languages for programming accelerating equipment to which everyone can contribute and cooperate without reference to the supplier.”