The national labs at Oak Ridge (ORNL), Argonne (ANL) and Lawrence Livermore (LLNL) are banding together for their next refresh of supercomputers. In late 2016 or early 2017, all three Department of Energy (DOE) centers are looking to deploy their first 100-plus petaflop systems, which will serve as precursors to their exascale machine further down the line. The labs will issue a request for proposal (RFP) later this year with the goal of awarding the work to two prime subcontractors.

The trio of lab partners, known as CORAL (Collaboration Oak Ridge Argonne Livermore), sent out a Request for Information (RFI) in December 2012 to gather information for the upcoming RFP. It's possible three separate RFPs will be issued, corresponding to systems hosted at each lab, but according to the RFI addendum, the DOE is "strongly considering" wrapping the multiple acquisitions under a single RFP.

The CORAL partnership between ORNL, ANL and LLNL to secure these pre-exascale machines mirrors the approach of their DOE siblings, NERSC, Los Alamos and Sandia National Labs to acquire their next round of supercomputers. In the latter case, those centers are teaming up to deploy two new machines (NERSC-8 and Trinity) before the end of 2015, about a year ahead of their CORAL counterparts. Because of the time difference and the somewhat different user bases, NERSC-8 and Trinity are almost certainly going to be sub-100-petaflop systems.

The CORAL supercomputers are initially spec'd at 100 to 300 petaflops, along with 5 to 10 petabytes of memory and 70 to 150 PB of storage. "The expectation is that the proposed 2016-2017 system will be roughly an order of magnitude less in time-to-solution than today’s systems at our facilities," states the RFI. If everything goes as planned, that means the top supercomputer at ORNL in four years will be about 10 times as powerful its current top machine, Titan, which currently delivers 24 peak petaflops and holds title to the most powerful computer on the planet.

Of course, the labs' focus on "time to solution" is centered around the traditional DOE application domains DOE like molecular dynamics, cosmology, CFD combustion, and others that map to the agency's Office of Science and NNSA missions. Since these are all Fortran and C/C++ codes, which employ mostly MPI and OpenMP to extract parallelism, the new platforms must be designed to support both legacy codes as well as any future frameworks for exascale computing.

Although the CORAL lab acquisitions have been combined, two distinct solutions will be chosen. One of them will be delivered as separate systems to both ORNL and ANL, while LLNL will choose one of two solutions for its own use. Theoretically that could mean that all three labs could deploy the same machine, but since the feds likes to spread the supercomputing love around, it most likely means two system vendors will get the opportunity to deliver these pre-exascale machines.

More than likely, we're talking about IBM and Cray as the primes here, although SGI could also make a reasonable case for a leading-edge supercomputer. None of these vendors have revealed platforms topping 100 petaflops yet. Cray's latest supercomputer, the XC30 maxes out at 100 petaflops, and even at that level of performance, would rely on GPUs or Intel coprocessors that are still under development. IBM is no doubt working on its successor to Blue Gene/Q. But whether Big Blue's exascale roadmap continues to follow that architecture, incorporates their Power server technology, or comes up with something entirely novel, remains to be seen.

To help foster some of this development, part of the CORAL effort will be to fund non-recurring engineering (NRE) costs associated with these pre-exascale supercomputers. The intent is to pour up to $100 million into these NRE activities, the money to be split between the two prime subcontractors. Some of this could certainly filter down to processor vendors, memory makers, and interconnect providers as well.

It's up to the bidding vendors to impress the labs on how best to apply the NRE funding, for example, better programmability, improving memory performance, adding embedded network controllers, maximizing data transfers between heterogeneous components, developing more efficient power management, and so on. Alternatively, the NRE could be directed at accelerating schedules, improving system cost, or TCO. The idea is to fund technologies or processes that the IT market would not be expected to deliver naturally.

Both the CORAL and NERSC-8/Trinity efforts are very much in the tradition of the "swim lanes" procurement approach -- encouraging the development of competing supercomputing architectures by various labs and vendors. The DOE has simplified the process somewhat by splitting the six leading centers into two teams, each of which will seed money into exascale research via their preferred choice of industry players.

Since these systems will pave the way for exascale technologies, there's a lot at stake here for the vendors. This isn't, however, just restricted to a few elite machines for a handful of labs. Petascale supercomputers will become increasingly commonplace during the second half of this decade, and they will be based on many of the same technologies that will drive exascale systems. Those companies tapped by the DOE to develop these next-generation supercomputers will be in a prime position to build not just the first exaflop-capable platforms, but also a whole array of HPC products for a much wider market.