President Donald Trump’s new federal budget request keeps the United States in the global race to build the first exascale supercomputer. It’s one of the few scientific areas in the proposed 2018 budget to come out ahead.
This budget accelerates the exascale timetable by two years, requiring delivery of the first system in 2021 instead of 2023. It also calls for a second exascale system, with a different architecture, in the following year.
An exascale system is the next great leap in computing power. Supercomputing marks its progress in FLOPS, floating-point operations per second. The first system capable of one trillion FLOPS was delivered in 1997. This teraflop system was followed in 2008 by a petascale system. One petaflop equals 1,000 teraflops, and an exaflop is 1,000 petaflops or one quintillion floating point operations per second.
Supercomputing is an international race. China’s Sunway TaihuLight, at 93 petaflops, is the world’s most powerful supercomputer, according to the Top500’s latest ranking, released in June.
Achieving exascale requires solving hard problems. The system needs to use as little power as possible, and the software has to be able to deal with multiple failures likely in a system with millions of computing cores. Programming environments have to enable massive parallelism.
To get this rolling, the Energy Department last week (15 June) announced about $260 million in exascale research awards to six tech vendors, with more funding opportunities on the way. But the Trump administration’s approach to federal R&D poses a problem.
The funding increases in exascale development are arguably coming from cuts elsewhere in federal R&D. The National Science Foundation, for instance, will see its budget reduced by $841 million or just over 11 percent under the FY18 budget. This leaves it with $6.65 billion.
The R&D cuts have been widely condemned by the scientific community. The American Geophysical Union has called them “destructive under-funding.”
The American Association for Advancement of Science said the Trump administration’s proposals will “devastate America’s science and technology enterprise.”
Delivering an exascale system by 2021 is contingent on Congress. Lawmakers will have to appropriate $200 to $300 million to build a supercomputer, and likely millions more for additional research.
Trump’s broader proposal to cut R&D faces resistance from lawmakers. Research universities and their corporate allies, especially in the health and pharmaceutical sectors, will argue that the United States is risking jobs and technological leadership with R&D cuts.
Why increase exascale development and then cut funding to the science that needs these systems? There are three reasons for this.
First, Trump likes to “win” and on fast timelines. Basic science research can take decades to yield fruit, but exascale development, in contrast, is moving quickly. The development of exascale is truly a contest pitting the United States against China, Japan and Europe.
China has said it will deliver an exascale system in 2020, the world’s first. But what constitutes “first” in exascale? Is it hitting a Top500 benchmark, or must such a system demonstrate an ability to run a scientific application at scale? Unfortunately, there are no Olympic-style judges in supercomputing to sort this out.
“A lot of people have their eyes on China, and they’re worried about China,” said Steve Conway, research vice president of Hyperion Research’s High-Performance Computing (HPC) group. “If they are worried about U.S. leadership, they better be concerned about Europe because in some respects Europe is a lot stronger than China,” he said. Conway points, in particular, to Europe’s HPC user base as mature and ready for supercomputing, unlike China.
Indeed, eight European nations, Germany, Portugal, France, Spain, Italy, Luxembourg, Netherlands and Belgium recently announced an agreement to develop an exascale supercomputer. It has been described as an “Airbus-sized” alliance. Airbus grew out of an European initiative with the goal of challenging Boeing.
Second, supercomputing is critical to national defense. It plays a vital role in weapons systems, encryption, missile defense and other capabilities. Exascale funding is in line with the Trump administration’s plan to seek a large increase in defense spending.
Scientists at the National Security Agency and Energy Department met late last year and agreed, in a report, that China “has attained near-peer status with the United States” in supercomputing. A “loss of leadership in HPC will severely compromise our national security,” they wrote.
The third reason is federal high-performance computing investment directly helps the tech sector maintain its global leadership.
The firms sharing the DOE’s recent $260 million exascale development award are Advanced Micro Devices, Cray, Hewlett Packard Enterprise, Intel, Nvidia and IBM. U.S. vendors account for nearly 50 percent of the systems on the Top500 global supercomputing ranking. The vendors will provide additional funding that amounts to at least 40 percent of the total project cost, said the Energy Department.
Federal investments in exascale helps vendors develop systems for commercial customers. “We do also believe that the work we do has to be also be viable for commercial products and building custom systems,” said Jim Sexton, IBM fellow and director of data centric systems, at last week’s Energy Department announcement.
The Energy Department Office of Science will get $4.43 billion in FY18, a cut of just over 16 percent from the current year, under the Trump plan. This budget, however, increases exascale system funding by just over $100 million next year, a 16 percent increase, according to an analysis by the Computing Research Association.
Pushing forward, HPC computing “will force new innovations in power, reliability, speed, software and software development tools and a host of other areas,” said Peter Harsha, director of government affairs for the Computing Research Association.
While money for exascale is increased, other programs in DOE’s Advanced Scientific Research are seeing double-digit decreases, said Harsha. Of concern to the computing research community is that some of program funding for mathematics and computer science research will also get moved to exascale programs, he said.
Conway estimates that it will cost about $5 billion in public and private investment to develop an exascale system.
The exascale program development has budgeted about $430 million so far, according to Paul Messina, the government research scientist who heads the Exascale Computing Project. The recent funding round is intended to accelerate exascale research “beyond what the vendor or computer manufacturer roadmaps currently have scheduled,” he said at the funding announcement.
IEEE-USA supports the exascale investments, but is encouraging Trump to broadly support basic research, said Brendan Godfrey, the chair of the R&D Policy Committee.“Exascale computing is critical to the economy and national security of the United States,” said Godfrey. “We commend President Trump for recognizing this, and for proposing aggressive investments in exascale computing research. Of course, no technology can be developed in isolation. Progress in this field will rely on innovations in many others. We hope President Trump matches his vision in exascale computing with similar support for all basic research,” he said.