University scientists face frustration due to limited computing power for artificial intelligence (AI) research, as revealed by a worldwide survey. The study, shared on arXiv on October 30, highlights that academics often lack access to advanced computing systems, impacting their ability to develop large language models (LLMs) and conduct AI research. Universities frequently struggle with acquiring powerful graphics processing units (GPUs), essential for AI model training but costly. Large tech firms, in contrast, possess higher budgets to acquire thousands of GPUs. Apoorv Khandelwal from Brown University and co-author of the study notes that while industry giants might have extensive GPU resources, academics may have only a handful, creating a vast gap between industry and academic capabilities. Khandelwal's team surveyed 50 scientists from 35 institutions, finding that 66% of respondents rated their satisfaction with available computing power as 3 or less out of 5, citing delays in GPU access and significant disparities globally, such as in the Middle East.

Limited access deters many from engaging in costly pre-training of LLMs. Co-author Ellie Pavlick stresses the importance of a competitive academic environment for long-term technological growth, contrasting it with the commercial pressure in industry research. Despite these limitations, researchers examined ways academics could maximize less powerful hardware by pre-training LLMs with 1 to 8 GPUs. This requires adopting efficient methods, which, while lengthier, allowed successful model training despite limited resources. Ji-Ung Lee from Saarland University in Germany finds this approach promising, noting that even small companies face similar access challenges. This study underscores the importance of addressing resource disparities to enhance AI research in academia.