President Barack Obama-who during his January State of Union speech highlighted supercomputing as a way for the United States to maintain its economic edge-has requested additional funding for the development of exascale supercomputers in the federal government’s 2012 budget.
The proposed 2012 budget included a request for $126 million for exascale computing for the Department of Energy. If Congress approves the budget, the funds will be split between the department’s Office of Science for $91 million, and the National Nuclear Security Administration for $36 million. In comparison, the 2011 budget allocated $24 million for “extreme scale” computing.
The requested exascale funding is part of the DOE’s total budget of $465 million for “advanced” computing. The proposed 2012 budget reflects a 21 percent increase over the 2010 budget.
Exascale supercomputers are expected to be 1,000 times faster than existing petaflop supercomputers. Useful for modeling and simulating complex situations such as virtual nuclear weapons testing, nuclear fusion research, or weather forecasting, exascale supercomputers would have the speed and computing power required for complex workloads. Applications such as modeling regional climate change or simulating the effects of a new drug on the human heart require computers with the capacity to rapidly process enormous volumes of data. The most advanced applications are beyond the capability of today’s commercially available computers.
These exascale systems are expected to be available by 2021, according to several experts.
In his 2011 State of the Union speech in January, the president noted that the United States no longer had the fastest supercomputer in the world. The Tianhe-1 machine at the National Center for Supercomputing in Tianjin, China, reported sustained computing power of 2.507 petaflops, which is significantly faster than the U.S. DOE’s Cray XT5 Jaguar in its Oak Ridge, Tenn., facility, with a computing power of 1.75 petaflops.
While Congress considers the proposed budget, the federal government is going ahead with its supercomputing plans. IBM announced earlier this month that it would be deploying Mira, a 10-petaflop supercomputer, to the DOE’s Argonne National Lab in Illinois. A Blue Gene/Q supercomputer capable of 10 quadrillion calculations per second, Mira is expected to be operational in 2012, IBM said.
Mira will be a significant upgrade for the Argonne, which currently has Intrepid, another IBM Blue Gene supercomputer, which can perform 557 teraflops (557 trillion calculations) a second, IBM said. Mira will be 20 times faster than Intrepid, said IBM.
Mira’s computing power will come from more than 750,000 IBM PowerPC A2 1.6 GHz 16-way SMP A2 processors and be capable of 40G bps throughput, according to IBM.
IBM is currently working on Sequoia, a 20-petaflop supercomputer for the DOE’s Lawrence Livermore National Lab. Sequoia is also expected to be operational in 2012. With 96 refrigerator-sized racks containing a combined 1.6 PB of memory feeding 1.6 million IBM Power processor cores, Sequoia will handle analysis of the U.S. nuclear stockpile, IBM said.
IBM developed the Blue Gene architecture, which Sequoia and Mira are based on, in 1993. Two of its supercomputers are currently among the Top 10 most most powerful supercomputers in the world, with a total of 22 systemsin the Top 100 list. IBM is also building the 10 petaflop Blue Waters system for the University of Illinois at Urbana-Champaign’s National Center for Supercomputing Applications.
As data volumes continue to explode, the ability to compute in exabytes is becoming increasingly necessary. Monthly Internet data flow is estimated to be 21 exabytes. However, the challenge facing researchers working on exascale computing is not CPU processing, but rather that of power management, according to Peter Kogge, professor of computer science and engineering at the University of Notre Dame.
The development of exascale is liable to hit a “power wall,” Kogge explains, because the amount of power required by a machine capable of that kind of computing power would be “absurd.” Scaling the 10-petaflop Blue Waters system 100 times to reach exaflop processing will require 1.5 gigawatts of power to run it, which is greater than 0.1 percent of the total U.S. power grid, Krogge said in the IEEE Spectrum magazine. The current architecture and technology needs to be rethought for exascale computing, to bring “voracious power requirements down to manageable levels,” he said.
Kogge’s analysis was the result of a challenge posed by the U.S. Defense Advanced Research Projects Agency in 2007 to determine the kind of technologies needed by 2015 to build an exascale supercomputer.