超並列重力多体問題専用計算機GRAPE-4  [in Japanese] GRAPE - 4 : A massively - parallel special - purpose computer system for astrophysical many - body simulations  [in Japanese]

Access this Article

Search this Article

Author(s)

    • 泰地 真弘人 Taiji Makoto
    • 東京大学教養学部宇宙地球科学教室 Department of Earth Science and Astronomy, College of Arts and Sciences, University of Tokyo
    • 牧野 淳一郎 Makino Junichiro
    • 東京大学教養学部情報図形科学教室 Department of Information Science and Graphics, College of Arts and Sciences, University of Tokyo
    • 戎崎俊一 Ebisuzaki Toshikazu
    • 東京大学教養学部宇宙地球科学教室 Department of Earth Science and Astronomy, College of Arts and Sciences, University of Tokyo

Abstract

我々は約1Tflopsの性能をもつ超並列重力多体問題専用計算機GRAPE?4の開発を進めている。GRAPE?4のために粒子間の重力相互作用を高遠に計算する「HARP (rmite AcceleratoR Pi) chip」を開発した。HARP chipでは仮想マルチパイプラインアーキテクチュアを考案し、入力データ転送率を半減する工夫をした。HARP chipは1クロックあたり20演算相当をおこない、30MHz動作時に600Mflopsの性能をもつ。約2000個のHARP chipをマルチチップモジュール技術を用いて集積し、1.2Tflopsのピーク性能をえる。20万体以上を扱うときには実効性能で1Tflopsを越える。We are developing GRAPE-4, a massively-parallel special-purpose computer for astrophysical N-body simulations. We have developed an application-specific LSI, the HARP (Hermite AcceleratoR Pipe) chip. The HARP chip calculates the gravitational interaction between particles. We invented the "Virtual Multiple Pipeline" architecture for HARP chip, which decrease the input data transfer rate. The HARP chip performs about 20 floating point operations per clock cycle and works at 30 MHz in the worst case. Therefore, the performance of the HARP chip exceeds 600 Mflops. GRAPE-4 will consist of about 2000 HARP chips using multi-chip modules. The sustained performance will exceed 1 Tflops when the number of particles exceeds 2×10^5.

We are developing GRAPE-4, a massively-parallel special-purpose computer for astrophysical N-body simulations. We have developed an application-specific LSI, the HARP (Hermite AcceleratoR Pipe) chip. The HARP chip calculates the gravitational interaction between particles. We invented the "Virtual Multiple Pipeline" architecture for HARP chip, which decrease the input data transfer rate. The HARP chip performs about 20 floating point operations per clock cycle and works at 30 MHz in the worst case. Therefore, the performance of the HARP chip exceeds 600 Mflops. GRAPE-4 will consist of about 2000 HARP chips using multi-chip modules. The sustained performance will exceed 1 Tflops when the number of particles exceeds 2×10^5.

Journal

  • IPSJ SIG Notes

    IPSJ SIG Notes 1993(72(1993-HPC-048)), 33-40, 1993-08-19

    Information Processing Society of Japan (IPSJ)

Codes

  • NII Article ID (NAID)
    110002932103
  • NII NACSIS-CAT ID (NCID)
    AN10463942
  • Text Lang
    JPN
  • Article Type
    Technical Report
  • Data Source
    NII-ELS  IPSJ 
Page Top