In what could be called a supercomputing arms race, China last year beat the U.S. by building the fastest computer in the world. The U.S. wants that title back.
The U.S. Department of Energy is allocating a total of US$425 million in research and development funds with the goal to ultimately build supercomputers about 20 to 40 times faster than the fastest systems today.
Of that, $325 million is being awarded to IBM and Nvidia to build supercomputers that will be five to seven times faster than existing systems in the U.S. The supercomputers will be installed at the DOE's Oak Ridge and Lawrence Livermore National Laboratories. It isn't clear when the supercomputers will be installed, but it could be within the next two years.
The systems will be based on CPUs from IBM and GPUs from Nvidia, which will help develop the supercomputers. IBM makes the general-purpose Power chip for servers and mainframes, while Nvidia's Tesla GPUs specialize in supercomputing calculations.
Supercomputers are central to the U.S.'s economic, natural security and environmental agenda. Supercomputers drive the space program, simulate economic models and predict weather. Faster supercomputers strengthen the country's general knowledge base and provide an edge in scientific and high-tech research over other countries.
The fastest supercomputer in the U.S. is called Titan and is used for research in areas including bioscience, climate studies, energy and space. Titan is being used for research in combustion engines, which aligns with the U.S. goal of improving fuel efficiency in cars and reducing oil imports.
China, Japan and the U.S. have been in a race to produce the world's fastest computers, exchanging top spots over the last few years. The world's fastest supercomputer today is China's Tianhe-2, which delivers sustained performance of 33.86 petaflops, scaling to 54.9 petaflops in bursts. Titan, which is installed at Oak Ridge National Laboratory in Tennessee, is in second place, delivering sustained performance of 17.59 petaflops, scaling to 27.1 petaflops. (A petaflop is equivalent to 1,000 trillion floating-point operations per second.)
The new DOE-funded supercomputers will surpass 100 petaflops of performance, which could bring a U.S.-made machine back to the top spot. The supercomputer at ORNL will be called Summit and deliver 150 to 300 petaflops of performance. The LLNL supercomputer will be called Sierra and deliver 100 petaflops of performance. It will replace a system called Sequoia, which was installed at LLNL in 2011.
The supercomputers are being developed as part of a DOE program called CORAL (Collaboration of Oak Ridge, Argonne, and Lawrence Livermore).
Funds are also being allocated by the DOE to solve the power-related challenges of making supercomputers 20 to 40 times faster than today's fastest systems. A major supercomputing problem has been trying to scale performance while reducing power consumption. DOE is funding IBM, Cray, Intel, Nvidia and Advanced Micro Devices as part of a program called "FastForward" to develop chips, memory, components and systems to address that problem.