Scientists working with the Department of Veterans Affairs' Million Veteran Program ran into a problem that faster supercomputers alone could not fix. Ravi Madduri's team at Argonne National Laboratory ran genome-wide association work on Oak Ridge's Summit system, but as the cohort grew, the matrix needed for each analysis no longer fit in CPU memory and the job stalled. When they opened up the code, they found the bottleneck was not raw speed but software: nearly all the time was spent multiplying large grids of numbers, an operation GPUs are built to handle.
Led by Argonne's Alex Rodriguez, the team rewrote that part of the widely used SAIGE genetics package for GPUs, creating SAIGE-GPU. The rewritten version turned a computation that would have taken roughly 481,000 CPU-core hours into a tractable run. Madduri points out that the same GPU architecture accelerating AI training is also foundational for population genomics.
With the GPU version, the Million Veteran Program consortium analyzed 2,068 traits across 635,969 veterans and flagged 13,672 regions of the genome linked to one or more conditions. About 1,600 of those associations appeared only because nearly a third of participants were not of European ancestry; each ancestry group had to be analyzed separately and the results combined. The consortium reported the results in Science and the computational work in Bioinformatics.
The two sources for this article are identical republished versions of the same text, so they agree on every point; no differences between them need to be reconciled.