Gloger.es@XEROX.ARPA (07/17/84)
The article quoted says film image production needs "raw computational speed," and suggests that this means a Cray-class computer. Could you not get comparable or better total computational speed at lower cost by running a roomful of microcomputers? Or is a Cray really the cheapest way to buy bulk floating point computation?
dave@rlgvax.UUCP (Dave Maxey) (08/10/84)
>The article quoted says film image production needs "raw computational >speed," and suggests that this means a Cray-class computer. Could you >not get comparable or better total computational speed at lower cost by >running a roomful of microcomputers? Or is a Cray really the cheapest >way to buy bulk floating point computation? Right now? Well, it's either that or the Cyber 205 (from CDC), or the supercomputer from Hitachi, or something else like it. Pipelines or vector pipelines as they are sometimes called are the best production machines on the market right now for raw computation speed. That's where you'll get the MegaFlops (Millions of Floating point OPerations per Second), or even GigaFlops. In the near future, data flow machines will probably take over, but there's nothing in the market right now that can beat the current supercomputers. In fact, I haven't even heard of ANY data flow machine on the market. I'm sure there must be some, they just haven't reached notoriety. A data flow machine is several processors linked together, each sharing the load of the process being worked on. This is a vast oversimplification of what it really means by data flow, but it should suffice for argument. :-) Actually, I could add that the idea is to be able to construct a task as a collection of subtasks, many of which can be executed in parallel. It requires a completely new approach to computing and programming. You are no longer thinking of things in just terms of sequential processing. You need new computer languages and compilers and machine instruction sets. The reason why a roomful of micros won't equal a supercomputer, is that you have to be able to get them to cooperate on the task. Doing that and expecting it to work for more than one task is more trouble than it could possibly be worth. You'd have to hand tailor each problem. And we're talking about hundreds or thousands of micros for one task, not just a roomful (I'm not just talking graphics here, but also things like seismology research, weather forecasting, etc.). It may be that you can get something like that to work for graphics, but for really high resolution, several colors, several gray scales, three dimensional rotating graphics in real time, with multiple light sources, etc. you need all the processing power you can get in one box. - Dave Maxey (alias tbm) {seismo,mcnc,brl-bmd,allegra}!rlgvax!dave
ken@turtlevax.UUCP (Ken Turkowski) (08/12/84)
The LINKS machine (an array of 64 Z8000'a) has been used in Japan to
make some impressive ray-traced animations. Yochiro Kawaguchi seems to
have used it the most, and presented a video at this year's and last
year's SIGGRAPH.
--
Ken Turkowski @ CADLINC, Palo Alto, CA
UUCP: {amd,decwrl,dual,flairvax,nsc}!turtlevax!ken
ARPA: decwrl!turtlevax!ken@BERKELEY