Research Computing expanding GPU capabilities to meet researchers’ AI needs

As computational research increasingly involves artificial intelligence and machine learning methods, ITaP Research Computing is adding GPU nodes to its community clusters to better serve researchers who use these tools. 

Research Computing recently added 24 new Nvidia A100 GPUs in the Gilbreth community cluster, 16 new AMD MI50 GPUs in the Bell cluster, and 8 new A100 GPUs in the Geddes composable platform. In addition, MI100 GPUs currently on loan from AMD are available to Purdue PIs for benchmarking.

This addition brings the total of centrally operated GPUs at Purdue to 158, with a total performance of 2.1 single-precision PetaFLOPs. These new GPUs represent a 45% increase in the number of GPUs and a 70% increase in the number of single precision FLOPs to support AI and machine learning research.

“The President’s Council of Advisors on Science and Technology has recommended a tenfold increase in federal investment in AI R&D funding over the next 10 years. To make sure that our Purdue faculty have the facilities on hand to be competitive with this revolution in AI research, our next several years’ plans in Research Computing reflect an increase in the annual investment in HPC resources designed for AI,” says Preston Smith, executive director of Research Computing. “This will bring to campus a regular refresh of current-generation GPUs, providing fast storage to feed AI models, and consistently curated software stacks.”

Community cluster supercomputing makes more computing power available for Purdue researchers than faculty and campus units could individually afford, and provides this crucial service at the lowest cost to the institution. ITaP Research Computing installs, administers and maintains the community clusters, providing reliability, security, software installation and expert user support.

Since 2004, the community cluster model has relieved Purdue researchers from needing to deploy and operate their own CPU-based clusters, and the efficiency and facilities challenges that comes with them. “Expanding the pool of GPUs in the community cluster program build AI-optimized resources is a natural extension of our proven business model.”, notes Smith.

“Rather than buying our own hardware to run in our lab, we’ve become heavy users of our A100 node as we rush to a conference deadline,” says Saurabh Bagchi, professor of electrical and computer engineering and computer science, who recently purchased a GPU node on Gilbreth. “Amiya [Maji, senior computational scientist for Research Computing] was prompt and helpful in getting one of our researchers, who was new to the cluster, up and running on the node.”

Purdue’s community cluster program began with just four faculty partners. Today the program has more than 200 active faculty partners from all three Purdue campuses, all of Purdue’s primary colleges and schools and 60 different departments. In 2020, ITaP Research Computing delivered 378 million computational hours to Purdue researchers, who account for 55% of Purdue’s FY20 sponsored research spending.

To learn more about GPUs or other Research Computing resources, contact Smith, psmith@purdue.edu or 49-49729.

Writer:  Adrienne Miller, science and technology writer, Information Technology at Purdue (ITaP), 765-496-8204, mill2027@purdue.edu.