The United States Department of Energy has announced a plan to field an exascale system by 2022, but says in order to meet this objective it will require an investment of $1 billion to $1.4 billion for targeted research and development. The DOE’s June 2013 “Exascale Strategy” report to Congress was recently obtained by FierceGovernmentIT.
The report makes it clear that exascale systems, one-hundred to one-thousand times faster than today’s petascale supercomputers, are needed to maintain a competitive advantage in both the science and the security domain. The DOE notes that exascale computing will be essential to the processing of future datasets in areas like combustion, climate and astrophysics and claims that there is “significant leverage in addressing the challenges of large scale simulations and large scale data analysis together.”
But before practical exascale machines can become a reality, there are several pretty major obstacles that need to be addressed. Among these are the energy issue; system balance and the memory wall; resiliency and coping with run-time errors; and exploiting massive parallelism. All of these issues require focused research and development.
Reducing power requirements is one of the foremost objectives of any exascale endeavor. The report points out that an exascale supercomputer built with current technology would consume almost a gigawatt of power, approximately half the output of Hoover Dam. With a standard technology progression over the next decade, experts estimate that an exascale supercomputer could be constructed with power requirements in the 200 megawatt range at an estimated cost of $200-$300 million per year. Whether funding bodies will be willing to spend this much money remains to be seen, but the DOE would like to see that power requirement cut by a factor of 10, down to 20 megawatt neighborhood where current best-in-class systems reside.
As a point of comparison, the largest US supercomputer, Titan, installed at Oak Ridge National Laboratory, requires 8.2 MW to reach 17.59 petaflops. The world’s fastest system, China’s 33.86 petaflop Tianhe-2, has a peak power load of 17.8 MW, but that figure goes up to 24 MW when cooling is added.
The DOE report recommends five main areas of focus which add up to a comprehensive exascale roadmap with the goal of fielding such a system by the beginning of the next decade (circa 2022).
- Provide computational capabilities that are 50 to 100 times greater than today’s systems at DOE’s Leadership Computing Facilities.
- Have power requirements that are a factor of 10 below the 2010 industry projections for such systems which assumed incremental efficiency improvements.
- Execute simulations and data analysis applications that require advanced computing capabilities such as performing accurate full reactor core calculations, validating and improving combustion models for mixed combustion regimes with strong turbulence-chemistry interactions, designing enzymes for conversion of biomass, and incorporating more realistic decisions based on available energy sources into the energy grid.
- Provide the capacity and capability needed to analyze ever-growing data streams.
- Advance the state-of-art hardware and software information security capabilities.
The plan described in the report covers the research, development and engineering that is needed to achieve an exascale computing system by 2022, but the acquisition of such a system would be separate from this effort. The suggested approach is to continue fielding systems at intermediate stages of performance, for example 100 petaflops, 250 petaflops, 500 petaflops, and so on, up to exascale. Currently, the US invests between $180M to $200M annually to acquire and operate HPC machines through the NNSA Advanced Simulation and Computing (ASC) and Office of Science Advanced Scientific Computing Research (ASCR) programs.
The R&D required to prepare the way for an exascale supercomputer comes with a price tag of between one billion and 1.4 billion dollars, a figure arrived at by surveying key stakeholders in the computing industry. This is the cost to the DOE with an expectation that there will be some “cost-share contribution” from vendors and some software componentry development left to the software ecosystem to resolve. Responsibility for the program will be jointly shared by the DOE’s Office of Science and the National Nuclear Security Administration (NNSA).