�ٺ���Ƶ

Artificial intelligence is all well and good until it melts the computers. Just ask , vice chair for research in the . He recalls the day one of his radiology projects churned through so many X-rays that it burned out multiple hard drives on �ٺ���Ƶ Health’s old high-performance computer system. “We overworked it so much that the circuits actually started smoking,” Dr. Sodickson says.

That problem is long gone. Last year, �ٺ���Ƶ invested in two new, more powerful high-performance computing systems: Big Purple, which fuels AI projects like the lung cancer diagnosis algorithm of ; and Skynet for image-intensive radiology projects like Dr. Sodickson’s fastMRI collaboration with Facebook.

The two systems share physical space in a secure 5,500-square-foot data center with backup power and cooling systems in Piscataway, New Jersey. Both are configured to protect confidential patient data while crunching millions of data-hogging graphics. “A single MRI can have 65 million pixels of information, and our research involves tens of thousands of imaging studies, so you can see how the data add up,” says Yvonne Lui, MD, associate chair for artificial intelligence in the Department of Radiology.

Big Purple is capable of 18 quadrillion floating point operations per second, or petaflops, and Skynet is close behind at 17 petaflops. While these numbers can’t compete with the 200-petaflop speed of Summit, the Department of Energy’s top-ranked supercomputer, both high-performance computers are among the fastest at any academic medical center, says Martin Ossowski, PhD, director of high-performance computing at �ٺ���Ƶ.

Skynet (a tongue-in-cheek nod to the humanity-hating AI in The Terminator) has already reduced the time required to train a typical neural network to perform a complex task from days or weeks to hours.

Big Purple, meanwhile, runs AI algorithms at roughly twice the speed of its predecessor while also supporting the non-AI work of more than 700 researchers across the institution. “Big Purple and Skynet are a quantum leap forward for �ٺ���Ƶ,” says Dr. Sodickson. “Many AI projects we are doing now simply wouldn’t be possible without them.”