MICROELECTRONICS TECHNOLOGY ALERT, JANUARY 22, 1999
Even if you are not an astronomer,
you should take a good look at GRAPE (Gravity Pipe), a computer devised by researchers at University of Tokyo and Princeton's Institute for Advanced Study. It shows in high-end computing what we've been seeing more of at the low end: the value of dedication. If you hardwire a computer to perform a particular job, it will perform better and cheaper than anything you can get out of the general-purpose computers the industry has been built around.GRAPE is a family of special-purpose hardware that works much as a graphics accelerator does. The accelerator speeds graphics calculations on a workstation without changing the software. GRAPE, also an attached piece of hardware, acts as a Newtonian force accelerator. GRAPE itself is getting faster. GRAPE 1, in 1989, operated at a speed of 240 Mflops. The speed next year (depending on specific application) will be 1 Tflop for limited precision data path, 100 Tflops for arbitrary force law, and 200 Tflops for full precision data path.
Such add-on hardware could become important as computational physics emerges as a third branch of physics (joining theoretical and experimental physics), and similar use of computer simulations increases in other disciplines. Computers are now used to model a whole system directly with data analyses following virtual lab experiments, much as experimenters analyze data from lab experiments.
As impressive as increases in computer speed have been to most of us, it hasn't turned heads among physicists. They want much more. Today's chip contains two orders of magnitude more transistors than chips did 10 years ago, and clock speeds have gone up one order of magnitude, which should yield a speed increase of more than a factor of 1000. They aren't seeing it in ph ysics labs. Actual speed increase of a typical chip has been at most a factor of 100. The problem is the costs resulting from the growing complexity of a general-purpose chip. That's why they see a big payoff from a chip designed for only one specific purpose.
With GRAPE, in a large-scale gravitational N-body calculation (N is the number of particles) almost all instructions of the computer program are performed on a standard workstation. Only the gravitational force calculations, in the innermost loop, are replaced by a function call to the special purpose hardware. Force integration and particle pushing are all done on the host computer, and only the interparticle force calculations are done on GRAPE.
Since the intrinsic speed of GRAPE is faster by a factor of 10,000 than the host computer, this could be a problem. But the interparticle calculations require a computer processing power that scales with N squared while all other actions on the host scale only in proportion to N. Each doubling of the number of particles doubles the workload on GRAPE relative to that on the workstation. No matter how slow the workstation, it will keep up with GRAPE if N is big enough.
Look for more high-end dedication, just as we have seen low-end dedication, chips designed to perform just a few functions easily and cheaply in automobiles and appliances. We'll see more high-end supplementary computer augmentation as supercomputer capacity gets down to a price level single labs can afford. The big supercomputer centers must serve all sorts of applications, so they must be basically general purpose computers. But a group of astrophysicists, say, will only want to do a few things. They will want to turn their workstations into dedicated tools that will do these things better.
Details: Piet Hut, Institute for Advanced Study, Princeton, NJ 08540. Phone: 609-734-8075. Fax: 609-924-8399. E-mail: piet@ias.edu.
back