By Max A. Cherney and Stephen Nellis
TORONTO/SAN FRANCISCO, April 6 (Reuters) – A niche acquisition by Nvidia has raised concerns among artificial-intelligence and supercomputer specialists who see the move as a test of the biggest AI chip company’s commitment to maintaining a fair playing field for chip rivals and AI data center builders.
Nvidia last December announced it would acquire SchedMD, giving it control of open-source software called Slurm, which schedules computing tasks and is critical for training large language models that power chatbots such as Anthropic’s Claude. Slurm also runs on government supercomputers which help forecast the weather and develop nuclear weapons.
The Slurm software helps power about 60% of supercomputers worldwide, according to SchedMD. Some of the engineers and executives who use those systems fear that Nvidia will subtly favor itself, five people say, such as by writing software updates for its own chips before those of rivals like Advanced Micro Devices. Slurm is used to manage Nvidia chips operating in supercomputers or AI-focused data centers.
There also is a hope among some users that Nvidia, the world’s most valuable publicly traded company, will reinvigorate development, pouring some of its staggering resources into long-awaited updates of a system built years ago for government supercomputers and now spreading from national labs into frontier AI companies.
The conversations among AI industry executives and supercomputer experts, and the concerns about Nvidia’s actions, were previously unreported. In a statement last week, Nvidia said: “Customers everywhere benefit from our open source and free software. Slurm is open-source and we continue to provide enhancements for everyone.” When announcing the SchedMD acquisition, Nvidia also said it was committed to developing and widely distributing the “open-source, vendor-neutral software.”
NVIDIA’S COMMITMENT QUESTIONED
Nvidia has positioned the acquisition as a way to expand its investment in open-source technology that helps with AI’s development. Nvidia could help SchedMD users – particularly government labs – embrace newer techniques in AI alongside more traditional supercomputer work, said Addison Snell, CEO of chip consultancy Intersect360 Research.
At the same time, Snell said concern remained that Nvidia in the long run “could take what’s a common open-source tool and make it so that it works better or exclusively for its own parts, versus competing technologies such as those from Intel or AMD or any other AI processing company.”
Early tests could be how quickly Nvidia integrates new chips from AMD due later this year into Slurm’s computer code compared with how quickly Nvidia integrates the software with its own technologies such as InfiniBand networking chips, said one of the sources, an engineer who has worked with Slurm extensively on supercomputing systems.
Three of the sources who expressed concern with Nvidia’s SchedMD acquisition work in the AI industry and two have knowledge of supercomputer operations. All have used or developed systems that include non-Nvidia hardware.
Several other experts using SchedMD’s software did not express immediate concern about Nvidia’s acquisition, but said they were aware of such worries and that they were watching closely what the chip giant does with Slurm. Many of the supercomputer and AI industry people saw the acquisition as a test of Nvidia’s intentions.
SLURM’S WIDESPREAD ADOPTION
The software from SchedMD, pronounced, “sked-em-dee,” has been adopted by AI labs. Meta Platforms, French AI startup Mistral and Anthropic use it for some specific tasks, including elements of AI training. OpenAI uses another method based on technology developed by Alphabet’s Google, a spokesperson said.
Anthropic, Mistral and Meta did not respond to requests for comment.
The AI industry sources point to a previous acquisition by Nvidia as part of the basis for concern.
Nvidia bought Bright Computing in 2022. Bright Computing’s software is usable with non-Nvidia hardware, but it has been optimized for Nvidia, creating a performance penalty for users of other chips without additional work, the AI industry sources said.
Nvidia dismissed those claims, saying that Bright Computing technology supports “nearly any CPU or GPU-accelerated cluster,” referring to central and graphics processing units that form the backbone of data centers.
Anyone can take open-source code although it takes effort to produce fully working software.
“We encourage others to join us in contributing to the growing ecosystem of free and open source software,” Nvidia said, adding that it has a track record of continuing to provide free and improved offerings after acquiring open-source software firms.
Nvidia said it will “continue to offer open-source software support, training and development for Slurm to SchedMD’s hundreds of customers.”
(Reporting by Max A. Cherney in Toronto and Stephen Nellis in San Francisco; Editing by Jeffrey Dastin, Peter Henderson and Matthew Lewis)
Comments