preston@felix.UUCP (Preston Bannister) (11/16/87)
>>The ABAQ includes 3 "links", which are 10-megabit-per-second serial >>interfaces for talking to off-board transputers. >> >>Jack Lang, in his talk at the Atari press conference at Comdex, supposed a >>setup where workers each had their own transputer system on their desks, >>with all of them linked together and linked to a separate box containing >>many transputers. As an application's need for processing power increased, >>it could pull more transputers in. An intriguing concept -- throw the >>computer into high gear. >Apollo has this today - NCS - Network Computing System. Basically >a program can call subroutines on other machines on the network. >Even different computers (Cray's to PC's) on heterogeneous networks. I would expect that the inter-Transputer 'links', being implemented at the _hardware_ level, may be substantially faster than the procedure call & network I/O used by Apollo. If the granularity of the problem is large enough, then the difference in overhead wouldn't matter. For problems with smaller granularity, the low overhead inter-Transputer communication might allow you to distribute the work among a greater number of CPU's without getting swamped by overhead. Of course, it would be interesting (and valuable) for someone to actually prove this. Any takers? Atari? :-) -- Preston L. Bannister USENET : ucbvax!trwrb!felix!preston BIX : plb CompuServe : 71350,3505 GEnie : p.bannister