NSF Funds Artificial Intelligence Upgrade for Bridges

National Science Foundation Awards $1.8-Million for Upgrade to Expand Bridges’ “Deep Learning” Capabilities

Sept. 24, 2018

PSC’s Bridges supercomputer is being upgraded to provide the world’s most powerful AI servers to the national research community. The supplemental award of $1.8 million from the National Science Foundation funds acquisition of the resource, which features specialized graphics processing units (GPUs). “Bridges-DL’s” integration with the rest of Bridges and new staff positions will help researchers exploit the system’s full potential.

“PSC provides resources that let people go beyond ‘What can I do, given my local resources?’ to ‘How can I leverage data to make the next breakthrough?’” said Paola Buitrago, Director of Artificial Intelligence and Big Data at PSC and co-principal investigator for Bridges. “It’s breaking that barrier that unleashes creativity and accelerates discovery.”

Bridges-DL addresses the changing nature of research, which increasingly applies AI and GPU computing to drive transformational advances. It builds on Bridges’ strength in “converged scalable computing”—the combination of massive computational capacity with fast handling of Big Data, specialized components and ease of entry for scientists who need to scale up their computations but do not have a computational background. The new resource will provide CMU, Pitt, the Pittsburgh region and the nation with an extraordinary platform for AI and AI-enabled simulation. “DL” stands for “deep learning,” an AI technique that has repeatedly been shown to enable significant advances from information buried in large data. Deep learning is enabling advances such as helping radiologists increase the accuracy of diagnoses, driving the discovery of new materials for cheaper, more efficient energy production and selecting optimal crops to boost agricultural production. Bridges-DL will also be used for education and training, directly contributing to workforce development.

“Artificial intelligence is an integral part of the future of science,” said Rebecca W. Doerge, Glen de Vries Dean of the Mellon College of Science. “Bridges-DL will enhance the work being done at Carnegie Mellon, the University of Pittsburgh and other institutions in the region. I have no doubt that resources like Bridges-DL will significantly contribute to the scientific breakthroughs of tomorrow.”

“This award is a great testament to the continued relevance of high performance computing in general—especially in new areas such as AI—and to the significant footprint that PSC has in the national supercomputing arena,” said Rob Rutenbar, Senior Vice Chancellor for Research at Pitt.

PSC is a joint program of Carnegie Mellon University and the University of Pittsburgh.

Bridges-DL balances maximum capability and capacity, supporting the most complex deep-learning models with the highest accuracy and incorporating the largest data sets. Bridges-DL will make those transformative resources available—at no cost—to researchers across the country. The new resource, delivered by Hewlett Packard Enterprise (HPE), will add an NVIDIA DGX-2 enterprise AI research system and nine HPE Apollo 6500 Gen10 servers. The NVIDIA DGX-2 tightly integrates 16 NVIDIA Tesla “Volta” GPUs – the most powerful GPUs in the world—using the world’s highest-bandwidth on-node switch, the NVSwitch. When used as a single processing unit, the DGX-2 provides 2 petaflops of peak performance, 1,000 times faster than a high-end laptop.

The NVIDIA DGX-2 has about 107 times as many GPU “CUDA cores” and 128 times as much GPU memory as a high-end laptop. The HPE Apollo 6500 servers each have eight “Volta” GPUs internally connected by NVIDIA’s NVLink 2.0, providing substantial capacity to run many models concurrently. While Bridges-DL is primarily focused on AI, its GPUs can also accelerate certain other important applications. Bridges-DL will make possible research that cannot be done on other platforms, and it will allow other research to proceed up to 100 times faster than on users’ local computers.

Bridges-DL will be integrated with the existing Bridges system, providing a total of 17 petabytes of data storage, 29,036 CPU cores, 216 GPUs, 277 terabytes aggregate memory, and individual nodes with up to 12 terabytes of memory. Bridges, including Bridges-DL, is available at no charge for open research and education, and on a cost-recovery basis to industry.

“Pittsburgh, with the amazing things happening here in the life sciences, technology, and computer science, is the right place for Bridges-DL,” said Nick Nystrom, Interim Director of PSC and architect and principal investigator for Bridges. “We’re excited to bring this new opportunity to research in Pittsburgh and nationwide.”