The new Oak Ridge National Laboratories system will be used to help build an exascale computer running at 100 petaflops
The new system, called Summit, was unveiled last week ahead of this year’s Top500 list, which is due later this month.
It can process 200 quadrillion calculations per second, or 200 petaflops, compared with 93 petaflops for the Chinese system, according to IBM, which built Summit for the Oak Ridge National Laboratory (ORNL) in Tennessee.
IBM has been working on Summit and a twin system for Lawrence Livermore National Laboratory in California since 2014, when $325 million (£243m) was set aside by the Department of Energy for a programme called Coral.
The company said it has grown increasingly challenging to increase supercomputer speeds due to changes in the underlying chip technology, where rapid clock speed increases have given way to other kinds of advancements, including more efficient power usage and tightly integrated “hyperscale” data centres.
Summit uses new processor designs and fast storage and internal communications to deliver faster calculations, as well as making use of artificial intelligence to make better choices about what kinds of calculations are suitable for a particular application.
The system is to be used in areas such as designing chemical formulae and new materials, cancer and fusion energy research, astrophysics and climate modelling.
The Lawrence Livermore system, mainly funded by the DOE’s National Nuclear Security Administration, is more focused on military uses, including research into nuclear weapons reliability.
IBM says Summit will be used to determine what would be needed to build an exascale system, one that can process 100 petaflops.
Aside from raw speed increases, supercomputers must be tuned in order to run real-world applications at an unprecedented scale, the company said. Summit was built with 30 applications in mind.
The new system is made up of 4,608 interconnected computer nodes housed in cabinets that are cooled by pumps that run 4,000 gallons of water per minute through the system.
It takes up an eighth of an acre and at peak consumes about 15 megawatts, enough to power more than 7,000 homes.
Each node runs on two IBM Power9 chips that run at 3.1GHz, with each chip having 22 processing cores. They’re linked to six Nvidia Tesla V100 graphics chips, graphics chips being particularly well suited to supercomputing and AI tasks.
Each node has 1.6 terabytes of memory and can save data at 2.2 terabytes per second on a 250-petabyte storage system.
At Summit’s launch event earlier this month, ORNL laboratory head Dr Thomas Zacharia said a comparative genomics code was run on Summit while it was still under construction.
Speaking at the event, US Secretary of Energy Rick Perry said supercomputing was a “competition” that it was important for the US to win.
“This is about changing the world,” he said, according to the BBC.
As of last year’s list, China owned 202 of the systems on the Top500, compared to the US’ 143.
The US’ previous top supercomputer was fifth on the list.
Put your knowledge of artificial intelligence (AI) to the test. Try our quiz!