A survey conducted among scholars from numerous global institutions indicates that many university researchers feel frustrated by the limited computing resources available for artificial intelligence (AI) research.
According to reports, the inaccessibility of cutting-edge computing systems may hinder their ability to develop large language models and conduct other AI research.
Specifically, researchers in academia often lack sufficient access to powerful graphics processing units (GPUs), which are essential for training AI models and can cost thousands of dollars. In contrast, researchers at large tech companies typically have higher budgets to invest in GPUs.
“Every additional GPU increases computing power,” stated Apoorv Khandelwal, a computer scientist at Brown University and one of the study's authors. “Industry giants may have thousands of GPUs, while academia may only have a few.”
Stella Biderman, Executive Director of the non-profit AI research organization EleutherAI, remarked, “The gap between academia and industry is significant, but it should be much smaller. Research on this disparity is very important.”
To assess the computing resources available to academia, Khandelwal and colleagues surveyed 50 researchers from 35 institutions. Among respondents, 66% rated their satisfaction with computing power at 3 or lower on a scale of 5, indicating widespread dissatisfaction.
Universities have varying setups for GPU access. Some may have a central computing cluster shared by departments and students, where researchers can apply for GPU time. Other institutions may purchase machines for direct use by lab members.
Some researchers reported having to wait days to access GPUs, noting that wait times are particularly long around project deadlines.
The findings also underscore global disparities in access. One respondent mentioned the difficulty of finding GPUs in the Middle East. Only 10% of respondents indicated they had access to NVIDIA's H100 GPU, a powerful chip designed for AI research.
This barrier significantly complicates the pre-training process, which involves inputting vast datasets into large language models. “Due to high costs, most scholars dare not venture into pre-training research,” Khandelwal noted, emphasizing that the shortage of computing power could severely constrain future developments in this field.
“For long-term progress, having a healthy and competitive academic research environment is crucial,” said Ellie Pavlick, a scholar in computer science and linguistics at Brown University and another author of the study. “In contrast, industry research often faces significant commercial pressures, which can sometimes lead to a rush to results, reducing exploration of unknown areas.”
The team also explored how researchers can utilize limited computing resources more efficiently. They calculated the time required to pre-train multiple large language models in low-resource hardware environments using 1 to 8 GPUs. Despite facing resource constraints, researchers successfully trained multiple models, although the process took longer and required more efficient methods.