Science-driven system architecture: A new process for leadership class computing

Over the past several years, computational scientists have observed a frustrating trend of stagnating application performance despite dramatic increases in peak performance of high performance computers. In 2002, researchers at Lawrence Berkeley National Laboratory, Argonne National Laboratory, and IBM proposed a new process to reverse this situation [1]. This strategy is based on new types of development partnerships with computer vendors based on the concept of science-driven computer system design. This strategy will engage applications scientists well before an architecture is available for commercialization. The process is already producing results, and has further potential for dramatically improving system efficiency. This paper documents the progress to date and the potential for future benefits. An example of this process is discussed, using IBM Power architecture with a computer architecture design that can lead to a sustained performance of 50 to 100 Tflo p/s on a broad spectrum of applications in 2006 for a reasonable cost. This partnership will establish a collaborative approach to modifying computer architecture to enable heretofore unrealized achievements in computer capability-limited fields such as nanoscience, combustion modeling, fusion, climate modeling, and astrophysics