Energy Efficiency in Research Data Centers

Research computing and data facilities today face unprecedented opportunities alongside complicated challenges. Opportunities stem from the explosion of high performance computing applications leveraging artificial intelligence (AI), machine learning (ML) and deep learning, promising groundbreaking discoveries and new ways of doing business. But AI and the large language models (LLMs) now making their way into research computing come with a cost. AI applications need to run fast and work best on graphic processing units (GPUs), which are built for multitasking. GPUs have thousands of compute cores [1] and are designed to run complex workloads in parallel. The enormous growth of AI applications is due to the compute power of GPUs and the availability of very large data sets. Both CPU and GPU chips continue to pack more transistors into them and require more power. Future CPUs will require more than 500W of power, while GPUs will exceed a kilowatt, and both will require direct to chip liquid cooling to manage the heat load. Moreover, training AI models takes time and effort; often several tries are needed to create a good model. To complicate matters even more, AI and the use of GPUs comes as the world struggles to mitigate the impacts of climate change.

Data centers must balance the need for more capable AI-enabled infrastructure with the need to manage the environmental impact of research computing. Innovative new architectures have been developed to address the problem, but technologists and researchers need access and education to use them effectively. This paper highlights the main avenues to energy savings in information technology and computing infrastructure and suggests actions that can be taken to create research data centers capable of handling the most demanding problems, while meeting energy efficiency requirements at all stages of computing workflows and reducing their carbon footprint.

DOI: 10.13140/RG.2.2.25923.41767