|
Why do we need vectors? According to Mohr, there is a lot of legacy code out there that is vectorized and is unparallelisable. Examples are NASTRAN and Dyna. Vector processors have a very high single-processor performance and, very important have a very high memory bandwidth. Because in the US all effort has been put in scalar massively parallel processing, there is no real US solution for all the older C90/T90 systems. Requirements for such systems are high sustained application performance and high memory bandwidth. Only PVP systems can deliver that. Mohr estimates this market to be rather large: 0.5 - 1. billion Euro. Who will enter this battlefield, which will concentrate first in neutral Europe? Mohr sees three major vendors, Cray, NEC, Fujitsu and two newcomers, Tera and SRC, entering. SGI/Cray has the SV-1 and its follow-on the SV-2 as follow on to the C90/T90. Although these systems are cheaper to buy and to maintain, they do not offer much performance improvement. NEC SX-5, according to Mohr, is big, it is fast and it is vector. The Fujitsu VP series is parallel, vector and distributed. Newcomer Tera comes with a new architecture, the MTA, which tries to hide latency instead of beat it. SRC, former Cray computer enter with the SRC-6, in which they make a large fast memory. Although this is a nice spectrum of solutions, as long as you are outside the USA, and perhaps Japan, there is nut much choice within the USA. Or, perhaps some global companies have already moved their computation to countries where there is more choice? A company like Chrysler/Mercedes has enough options... Jeffry Mohr is an RCI advisor. Check in at the RCI web site for more information.
Ad Emmen
|
 |