e50214@kuduts.kudpc.kyoto-u.ac.jp (toshio nishigaki [e50214]) (07/05/90)
I am thinking of building a computer system for the exclusive use of MD
(moleculer dynamics) simulation (presentry in my mind but someday in LSI chip).
MD culculation is usually done on VP (vector processor) machine in
order to reduce run time. Although source code of MD culculation seems to be
easily vectorized, big loop of atom-atom interaction part will never be
completely vectorized on general-purpose computers. That's because VP is not
big enough to manage N atoms at one time, when N is 1000 or so. Besides, VP
machine is too expensive only to do MD simulations. I need no general-purpose
one to solve 1000 atoms system, but I want a strong VP of my own. This is my
basic idea.
Suppose there is a computer room, in which one teacher and 1000
students are sitting in front of their own PCs (personal computers). MD
simulation is done as follows. Teacher hands out 1000 data of atom coordinates
to his sutudents through LAN (local area network) and give the cue, then every
student starts cuculation on his own atom and give back the result to their
teacher. Teacher collect and reculculate them for the next stage. I mean, one
CPU for one atom.
How much does one CPU costs ? 32-bit CPU with CO-PROCESSOR costs more
than 5000 yen in JAPAN for powerful culculation and 8-bit CPU costs less than
1000 yen for economical culculation. Either will do. The former may be
considered in the class of SUN WS (work station) and the latter in the class of
PC of 10 years ago. (By 'CPU', I mean CPU chip or board here.)
Does anyone have already got this kind of machine ?
/*******************************************************************************
Toshio Nishigaki (E-mail : e50214@kuduts.kudpc.kypto-u.ac.jp)
Division of Theoretical Chemistry, Department of Chemistry,
Faculty of science, Kyoto Univ, JAPAN
*******************************************************************************/briscoe-duke@CS.YALE.EDU (Duke Briscoe) (07/05/90)
In article <946@kudpc.kyoto-u.ac.jp> e50214@kuduts.kudpc.kyoto-u.ac.jp (toshio nishigaki [e50214]) writes: > > I am thinking of building a computer system for the exclusive use of MD >(moleculer dynamics) simulation (presentry in my mind but someday in LSI chip). ... > Does anyone have already got this kind of machine ? > >/******************************************************************************* > Toshio Nishigaki (E-mail : e50214@kuduts.kudpc.kypto-u.ac.jp) > Division of Theoretical Chemistry, Department of Chemistry, > Faculty of science, Kyoto Univ, JAPAN >*******************************************************************************/ This sounds very similar to the problem of building a special purpose computer for predicting the stability of the planetary orbits in our solar system. A computer called the Digital Orrery was built to do this, and it was used to predict the orbit of Pluto hundreds of millions of years into the future, in order to find out if the orbit was stable or not. The computer was built using 1980 technology and was about 1 cubic foot in size, used 150 watts, but was 60 times faster than a VAX 11/780 and 1/3 the speed of a Cray 1 FOR THE SPECIFIC PROBLEM it was designed to solve. I've read that it was designed and built by six people working part-time for nine months, and it was relatively low cost compared to the alternatives, such as using a supercomputer. The calculation of several hundred million years of Pluto's orbit (to high precision) required the Orrery to run continuously for five months. There is a brief discussion of the use of specialized computers for numerical modelling on page 553 of vol. 32, Number 5, May 1989 Communications of the ACM. A reference that article gives for the Digital Orrery is IEEE Trans. Comp. C-34, 9 (Sept. 1985), pp. 822-831. This work was done by a group of people at MIT, and they wrote the two references I've given. Duke