GRAPE Project: A Decade of Special-Purpose Computers for Many-Body Simulations
Junichiro Makino, Masaki Koga, Atsushi Kawai, Toshiyuki Fukushige · Progress of Theoretical Physics Supplement · 2000
In this paper, we briefly overview the past history and future prospect of the GRAPE project to develop and use special-purpose computers for astrophysical many-body simulations. First we show that the “hardware efficiency” of general-purpose computers has been going down exponentially, and will continue to do so for the foreseeable future. Then, we describe the approach of building special-purpose computers as an alternative. With general-purpose design, we can use only a small fraction of available transistors on a chip, since most of the transistors are wasted in the control logic and the datapath between arithmetic units and the storage. With a special-purpose design, one can use virtually all transistors to implement arithmetic units, since little control logic is necessary and the datapath is fixed. This difference results in a huge difference in the price-performance, as is observed in GRAPE series hardwares. Of course, special-purpose computing is not a silver bullet. It has its limitations. we discuss these issues and possible alternatives. We also briefly describe the past of GRAPE project, and the status of GRAPE-6, which will be completed by the year 2001. It will provide the peak speed exceeding 100 Tflops, for the development cost of less than 4 M dollars.