For decades, the U.S. took for granted the doubling of supercomputing power every 10 years, roughly in line with Moore's Law. But once a petascale system was reached in 2008, it gradually became clear that the next leap -- a system 1,000 times more powerful -- would be difficult.
Initially, some believed such a system -- an exascale computer -- was possible in 10 years, or by 2018. But problems emerged. It took too much power, and it required new approaches to applications to utilize an almost unimaginable level of parallelism involving hundreds of millions of cores. Another problem to solve was the need for resilience, or an ability to continue to working around multiple ongoing hardware failures expected in a system of this size.
The new deadline is 2023, or 15 years after reaching petascale. An exaflop is a million trillion calculations per second (one quintillion). That is 1,000 times faster than a petaflop.
In other words, the government is aiming for a system that can solve science problems 50 times faster than the 20-petaflop systems now available.
"The U.S. faces serious and urgent economic, environmental, and national security challenges based on energy, climate, and growing security threats," wrote the U.S. Department of Energy, in a briefing document that was prepared for systems vendors. They met earlier this month with U.S. officials for an information session about the systems.
"High performance computing (HPC) is a requirement for addressing such challenges, and the need for the development of capable exascale computers has become critical for solving these problems," the government said.
This system, as imagined, will use between 20MW to 30MW of power, or roughly the output of a small power plant.
U.S. spending reflects this. The government will spend nearly $300 million this year on developing an exascale system, and has proposed spending a little more than that next year. The total cost for an exascale system is estimated at about $3 billion, according released DOE planning documents.
The U.S. is also in a race with China, Europe and Japan to build an exascale system.
Jack Dongarra, a computer scientist at the University of Tennessee and one of the people behind the Top 500 ranking of the world's most powerful supercomputers, said China has two projects aiming at 100 petaflops, and will probably soon make an announcement.
Steve Conway, an analyst at IDC, said the race to build an exascale system "is really up for grabs at this point." It's hard to tell who will get there first, what they are actually crossing the finish line with and "how useable it will be."