Creating Science-Driven Computer Architecture: The "Blue Planet" Proposal

In recent years scientific computing in America has been handicapped by its dependence on hardware that is designed and optimized for commercial applications. The performance of the recently completed Earth Simulator in Japan, which is five times faster than the fastest American supercomputer, dramatically exposed the seriousness of this problem. Typical scientific applications are now able to extract only 5 to 10 percent of the power of American supercomputers built from commercial web and data servers. By contrast, the design of the Earth Simulator makes 30 to 50 percent of its power accessible to the majority of types of scientific calculations.

Blue Planet was born here: At a two-day workshop in August 2002, a team of Argonne, Berkeley Lab, and IBM scientists developed the fundamental concepts of Virtual Vector Architecture (ViVA), potentially redefining supercomputing in America.

The Ultrascale Simulation for Science website provides information about the challenge posed by the Earth Simulator and the emerging U.S. response to that challenge.

Lawrence Berkeley and Argonne national laboratories, in close collaboration with IBM, have responded to the challenge with a proposal for a new program to bring into existence a new class of computational capability in the United States that is optimal for science. Our strategic proposal, "Creating Science-Driven Computer Architecture: A New Path to Scientific Leadership," envisions a new type of development partnership with computer vendors that goes beyond the mere evaluation of the offerings that those vendors are currently planning for the next decade.

 This comprehensive strategy includes development partnerships with multiple vendors, in which teams of scientific applications specialists and computer scientists will work with computer architects from major U.S. vendors to create hardware and software environments that will allow scientists to extract the maximum performance and capability from the hardware.

The cost of scientific supercomputing is also an issue of national strategic importance. The strategy we propose to implement will pursue at least three options:

1. At the highest cost per peak teraflop/s, the first option will involve custom components at all levels in an architecture known to be successful in scientific applications, parallel vector processing. The initial stages of this effort have been announced with the evaluation of a beta-test version of the Cray X1 at Oak Ridge National Laboratory.
2. At half this price and with the promise of sustainably high cost-effectiveness, the second option will involve commercial microprocessors in a new architecture that will be programmable in the same way as the first option, ViVA or Virtual Vector Architecture. IBM will partner with Lawrence Berkeley National Laboratory to implement early versions of this architecture and deliver Blue Planet, a 160 teraflop/s mature implementation in the second half of 2005.
3. At the lowest cost per peak teraflop/s, the third option will be based on "system-on-a-chip" architecture that is being explored most visibly in the IBM Blue Gene project. This architecture is arguably the most promising for reaching the petaflop/s goal of this proposal; however, its suitability for general scientific use has not yet been demonstrated. But at half the cost of option 2 and one quarter the cost of option 1, this is path is extremely cost-effective to pursue and provides the best long-term bet currently known to the scientific community. IBM will partner with Argonne National Laboratory to develop new expressions of this architecture and deliver a 180 teraflop/s implementation appropriate for general scientific exploitation in 2005.

Option 2 will provide twice the sustained capability of the Earth Simulator at half the cost. Option 3 will provide a new architecture family for scientific computing and one that makes a definitive step towards cost-effective petaflop/s computers with high sustained levels of performance.

In response to the High End Computing Revitalization Task Force (HECRTF) Outreach, the Berkeley Lab-Argonne proposal was rewritten in May 2003 with the participation of Lawrence Livermore, Oak Ridge, Pacific Northwest, Sandia, and Los Alamos national laboratories.