But the Barcelona Supercomputer Center is hoping to solve this problem using the low-power chip architecture that powers so many of the world’s smartphones and tablets. The center has inked a deal with chip maker NVIDIA to build a supercomputer using the company’s Tegra cellphone CPUs, which use the ARM architecture. These processors will take the place of the x86 Intel-compatible processors that power most of today’s supercomputers.
“ARM comes from the embedded world where you start with a power limit first,” said Sumit Gupta, manager of NVIDIA’s Tesla High Performance GPU Computing business unit. The Barcelona Supercomputer Center is showing its ARM-based design at the SC11 Conference this week in Seattle, and it aims to demonstrate a 4- to 10-fold improvement in energy efficiency over today’s most energy-efficient supercomputers by 2014.
Separately, the Barcelona Supercomputer Center has teamed up with Intel to create an Exascale Laboratory to focus on scalability of programming and runtime systems. Intel announced the relationship at the SC11 Conference. Intel also announced its “Knights Corner” co-processor, an x86-compatible floating-point processor capable of 1 teraflop double precision floating point performance.
The Barcelona ARM supercomputer will use a hybrid architecture that also includes NVIDIA’s CUDA graphics processors. GPUs are well suited to high-performance computing because graphics processing requires heavy floating-point number crunching. NVIDIA has been pushing the technology for supercomputing in recent years. Thirty-five supercomputers on the Top500 list of the world’s fastest computers use NVIDIA’s graphics processors as performance accelerators, including the number two, four, and five computers. Thirteen of the Green500 list of the world’s most energy-efficient supercomputers use NVIDIA GPU’s, including number five and six.
Once upon a time, the mightiest computers in the land were powered by special processors designed to crunch the huge amounts of data needed to do the sorts of things only supercomputers can do — model the climate, simulate nuclear explosions and sift through countless chemical compounds in search of new drugs. The high cost of those supercomputer chips drove the industry to instead string together thousands of off-the-shelf microprocessors.
The transition wasn’t easy, however. Shifting to commodity hardware came at the cost of developing a new software environment: new development tools and new or rewritten applications. The move to low-power processors will require a similar software makeover.
With that in mind, NVIDIA also announced a software development kit for supercomputer makers interested in NVIDIA’s hybrid ARM-GPU architecture. The kit includes a Tegra 3 CPU, a GPU and NVIDIA’s parallel programming toolkit. The Barcelona researchers “want to see how the applications scale with this architecture,” said Gupta. “[The project] gets the software ecosystem started.”
The Barcelona project isn’t the only one tapping embedded processors for supercomputers. Researchers at Lawrence Berkeley National Laboratory are building their own ultra-low-power processor using an embedded-processor-building toolkit from Tensilica. And the Barcelona project isn’t the only ultra-low-power supercomputer effort using NVIDIA’s architecture. NVIDIA is leading one of four projects under the DARPA Ubiquitous High Performance Computing program, which is encouraging development of prototype energy-efficient exascale supercomputers by 2018.
The DARPA UHPC projects are attacking the problem by going back to the drawing board on every aspect of supercomputing, from chips to applications. The Barcelona researchers, in contrast, are aiming to see how far they can go with off-the-shelf hardware.
Energy efficiency is also a concern for data center operators. With an eye toward addressing this problem, server makers are evaluating ARM processors, as well as Intel’s Atom processor. HP recently announced that it is developing a server using ultralow power ARM chips from startup Calxeda. “Our focus right now is on high-performance computing,” said NVIDIA’s Gupta. “I’m delighted if Calxeda and HP succeed because it sets up the market for us on the HPC side.”