Calculation at AI computing: Scientists have no access to powerful chips for their research

Calculation at AI computing: Scientists have no access to powerful chips for their research

Many university scientists are frustrated with the limited computing performance, which for their research in the field of Artificial intelligence (KI) is available, as a survey among academics at dozens of institutions worldwide shows.

The results 1 , which were published on the Preprint server Arxiv Most advanced computing systems lack. This could affect your ability, Large voice models (LLMS) to develop and carry out other AI research projects.

In particular, academic researchers sometimes do not have the resources to make powerful Graphics processors (GPUS) to be purchased-computer chips that are often used to train AI models and can cost several thousand dollars. In contrast, researchers have higher budgets in large technology companies and can spend more for GPUs. "Every GPU adds more power," says the co -author of the APOORV Khandelwal study, an computer scientist at Brown University in Providence, Rhode Island. "While these industrial giants may have thousands of GPUs, academics may only have a few."

"The gap between academic and industrial models is large, but could be much smaller," says Stella Biderman, managing director of Eleutherai, a non-profit AI research institute in Washington DC. Research on this inequality is "very important", she adds.

slow waiting times

In order to evaluate the available computing resources for academics, Khandelwal and his colleagues surveyed 50 scientists from 35 institutions. Of the respondents, 66% evaluated their satisfaction with their computing power with 3 or less on a scale of 5. "They are not satisfied at all," says Khandelwal.

The universities have different regulations for access to GPUs. Some could have a central compute cluster that is shared by departments and students, where researchers can request GPU time. Other institutions could buy machines that can be used directly by members of the laboratory.

Some scientists reported that they had to wait for days to get access to GPUs and noticed that the waiting times were particularly high (see "Calculation resource acceptance"). The results also illustrate global inequalities in access. For example, a respondent mentioned the difficulties of finding GPUs in the Middle East. Only 10% of the respondents stated that access to Nvidia’s H100 GPUS , powerful chips that were developed for AI research.

This barrier makes the process of pre-training-the feeding of large data records in LLMS-particularly challenging. "It is so expensive that most academics are not even considering doing science in pre-training," says Kaufenwal. He and his colleagues are of the opinion that academics offer a unique perspective in AI research and that a lack of access to computing power could restrict the research field.

"It is simply important to have a healthy, competitive academic research environment for long -term growth and long -term technological development," says co -author Ellie Pavlick, who studies computer science and linguistics at Brown University. "If you have research in industry, there is clear commercial pressure, which sometimes tempt you to use and explore less faster."

efficient methods

The researchers also examined how academics could better use less powerful computing resources. They calculate how much time would be required to train several LLMs with hardware with low resource consumption - between 1 and 8 GPUs. Despite these limited resources, the researchers managed to successfully train many of the models, although it took longer and they had to apply more efficient methods.

"We can actually use the GPUs that we have longer, and so we can compensate for some of the differences between what industry has," says Kaufwal.

"It is exciting to see that you can actually train a larger model than many people would take, even with limited arithmetic resources," says Ji -ung Lee, the Neuroexplicite models at the University of Saarland in Saarbrücken, Germany. He adds that future work could look at the experiences of industrial researchers in small companies that also fight access to arithmetic resources. "It is not the case that everyone who has access to unlimited computing power actually receives this," he says.

1. Khandelwal, A. et al. Preprint at arxiv https://doi.org/10.48550/arxiv.2410.23261 (2024).

Download references