Core Computation

Fire up the GPUs: UW–Madison, Morgridge project sparks next-level computing

A form of computing machinery that was once the province of hardcore video gamers — the graphic processing unit, or GPU — has recently taken the world of scientific research by storm.

Originally designed in the late 1990s with the capability of rendering 3D graphics, GPUs have been essential over the years to creating increasingly sophisticated and realistic visual effects.

While most of the research world has thought in terms of CPUs — or central processing units — as the lingua franca of computing power, GPUs are now emerging at the top of the rack for fields such as machine learning and scientific computing.

Anthony Gitter
Anthony Gitter

Morgridge Investigator Anthony Gitter, a UW–Madison associate professor of biostatistics and medical informatics, recognized the need early on in his machine learning projects related to protein engineering and drug discovery — projects that generate millions of data points. There were GPU-related tools available that could complete his team’s modeling experiments in days that would have taken months or years — if accomplished at all — with standard CPU-based computing.

But he also noticed, around 2018, a groundswell of DIY efforts across the UW–Madison campus related to GPUs.

“I saw a lot of my peers were trying to set up their own systems,” he recalls. “People were buying workstations that would have one GPU and sticking it under a desk for a grad student to run, then trying to figure out what hardware to buy, how to keep it maintained and what software to install.”

Gitter spotted an opportunity. Why not create a centralized resource and user community that could help support hundreds of varied GPU experiments, much like his Morgridge and UW–Madison colleagues have accomplished through the Center for High-Throughput Computing (CHTC)? That center successfully manages more than 300 unique projects a year, generating hundreds of millions of hours of computing time.

“We were able to bring along the documentation, build a user community and teach workshops on how to use the resource. We’ve just built this environment where we’re now really well set up for getting the most out of the technology.”

Anthony Gitter

The idea led, in 2019, to a successful grant request through UW 2020, a project supported by the Wisconsin Alumni Research Foundation (WARF) and designed to stimulate highly innovative projects that could potentially transform a field of study. Gitter assembled an investigative team that was co-directed by CHTC members and included faculty from more than a half-dozen data-intensive research fields.

The end product, in only a few short years, has been a remarkable return on investment. UW–Madison scientists now have free access to a pool of more than 100 GPUs, which are already being accessed by researchers from more than 40 UW–Madison academic departments. The GPU Lab is managed by CHTC with the expectation of incremental growth to meet research user needs.

The team built the resource slowly, adding on technology that specifically responded to feedback from the research community. After a series of investments, a second major investment by WARF and the Division of Information Technology (DoIT) in 2022 more than doubled campus GPU capacity.

“I think this investment was most successful because we didn’t plunk it all down on hardware,” Gitter says. “We were able to bring along the documentation, build a user community and teach workshops on how to use the resource. We’ve just built this environment where we’re now really well set up for getting the most out of the technology.”

Christina Koch
Christina Koch

Christina Koch, a research computing facilitator for CHTC, works closely with many of the scientists in the GPU space, and she says the resource has opened doors to a lot of unique projects. For example, it’s being employed by the Digital Livestock Laboratory that is using video data to track the health and nutrition of dairy herds. The Cryo-EM microscopy community is another major user.

Other examples include the Computational Materials Group in the College of Engineering, which is using machine learning to study the properties of metallic glasses. And chemistry Professor Xuhui Huang is using GPUs in machine learning experiments to predict interactions between chemicals and proteins.

“GPUs, as a technology, require a lot of energy, and generate a lot of heat and noise,” Koch says. “So I think that’s another good reason for people to try to partner with CHTC. We also have this user support focus at CHTC, which we hope is bringing the best value possible to campus and also incorporating researcher input.”

“I think it makes us very competitive with peer institutions, in terms of having access, and especially having it be free to campus researchers,” Koch adds. “And our curated approach is probably a little different than other campuses that might just offer a GPU farm.”