IBM supercomputer poised to be the world’s largest, fastest AI system
An IBM- designed US supercomputer unveiled on Friday is set to leapfrog Chinese competition to become the world’s most powerful for the first time in more than five years.
The machine, built by IBM for the Department of Energy’s Oak Ridge National Laboratory, has more than double the capacity of the current world leader, China’s Sunway TaihuLight.
But the real significance may lie in something other than its raw processing power: it uses a new computer architecture to handle huge amounts of data for artificial intelligence, rather than being limited to the kind of largescale modeling and simulations that such systems are normally used for.
Known as Summit, the supercomputer boasts 200 petaflops, or quadrillion floating point operations per second — the main measure for supercomputer capacity. That is about 80 times the processing power claimed for a Chinese system when that country first topped the world supercomputer rankings in 2010.
The Summit machine was designed with big data in mind, though on an even more huge scale than the specialized systems used by companies such as Google and Microsoft to handle the data needed to train artificial intelligence (AI) algorithms.
It breaks new ground among supercomputers as the first to be “designed to be the world’s largest and fastest AI system from the ground up,” said John Kelly, head of cognitive solutions and research at IBM. “It will be able to take on some of the biggest AI challenges the world has.”
Thomas Zacharia, director of Oak Ridge, said the system, which cost about $200 million, had already been used to carry out a calculation in an hour that would have taken 30 years on a desktop computer.
He added that plans for the system include feeding it all the medical records maintained by the US Department of Veterans’ Affairs — a significant trove of medical information that might yield new insights into how to treat common ailments among military veterans.
To improve the computer’s ability to handle large amounts of data, it was designed to bring processing power close to the system’s memory. It comprises more than 4,600 servers, each with direct access to a slice of memory and consisting of two of IBM’s Power processors and 6 GPUs from Nvidia, all linked together with lightning-fast interconnects.
The design removes a bottleneck that would otherwise have limited the potential of the GPUs, which act as accelerators in blazing through data, said Ashish Nadkarni, an analyst at IDC.
For IBM and Nvidia, the tech companies behind Summit, that also makes the machine a shop window for the latest AI technology. IBM plans to sell smaller versions of the system and eventually to offer access to one of the supercomputers as a service, said Mr. Kelly.
“The fact that you can get so much raw power out of [the IBM chips] should be worrying for Intel,” said Mr. Nadkarni. The technology could be used in industries that need to sift through very large amounts of data, like fraud detection in the financial industry and medical research, he added.
The length of time it takes to design and build such huge