SAN JOSE, Calif. – Engineers are exploring radical shifts in computing to deliver exascale-class systems by 2018. Their efforts recognize that today's best designs are too complex and power hungry to execute a quintillion instructions per second.
"If we don’t do this kind of work, exascale systems will consume 100 megawatts instead of the five MW today's top supercomputers use," said Richard Murphy, principal investigator overseeing a Defense Advanced Research Projects Agency program on the topic.
DARPA's Ubiquitous High Performance Computing program tasked four separate teams to pack a petaflop-class system into 57 kilowatt rack prototype computer by 2014. Such power-lean systems could be the building blocks for future exascale systems.
"We are trying to re-invent the future of computing to meet performance, power and programmability goals," said William Dally, chief scientist at Nvidia who is heading one of the DARPA teams.
Indeed, the complexity of programming supercomputers is another top challenge given today's petaflops systems typically pack tens of thousands of cores.
"The shift to multicore processor architectures really is stressing the existing programming model," said Murphy, a senior member of technical staff at Sandia National Labs. "In ten years, we will need a different way of thinking about how these machines work," he said.
Murphy's Sandia team is developing networking concepts for sending heavily threaded messages between systems as the basic units of work. It also aims to eliminate the traditional load-store approach to computation by keeping large data sets in memory. The effort is based on the observation that big systems use more resources moving data around than actually computing with it.
"What I think is we will end up with is a rethink of how we stage data from one place to another in a machine," said Murphy. "Most of the teams are looking at redesigning memory systems in some way, such as using deeper hierarchies, but our approach is to move computation as close to memory as possible," he said.
Intel and MIT researchers are heading the other two DARPA teams.