A bomb exploded - Cray and NEC cooperateMunich 02 March 2001 For more than 20 years in the high-performance computer and application market, I never observed such a co-operation between such competitors and between U.S. and Japanese vendors. A special highlight in their relationship is the anti-dumping offending of the former Cray Research Inc. What can be expected in this 10 year OEM agreement? Some of these topics are discussed in this very personal comment.
Cray
Cray was first based on its vector supercomputers of the Cray-1 line up to the C90 and T90. They are easy to use machines and very efficient - in the advisory service I supported the academic users in Lower Saxonian of the Cray 1M in Berlin at ZIB (Konrad Zuse Zentrum). But one day Cray made the same mistake as Control Data Corp., these machines sold very well and there was no need to change technology, improve the speed and keep up-to-date. In parallel the slow J90 series started in CMOS. But this technology was not used for the high-end machines. The successor SV1 as a Cray-compatible upgrade for J90s came late and did not perform as expected. The gap between 1.4+ GFlop/s an 10 GFlop/s peak performance per processor became greater. Although Cray reports that about 50 systems out of 100 computers are sold outside the U.S. The average selling price is well below the NEC, e.g. SV1 less US$ 1 mio list price, while the SX-5 costs more than 2 million US$. But as every head of the computer centre knows, these are starting prices and one ends with lesser costs.
Beneath that Cray Inc. sells the massively parallel computer T3E - which is now out of date - and the Tera MTA multithreading architecture. The first will be replaced by a cluster based on API NetWorks CS20, a 1U rack with two Alpha-processors. The MTA is always delayed, although one minor system is working at San Diego.
Now reading the press releases, Cray will distribute, sale, install and support the NEC SX-vector line. Perhaps the HNSX staff, the NEC subsidiary in the U.S. is big enough to take this work, but it is really a challenging task. The cooperation will open new chances and earnings for Cray in its home market.
But will Cray Inc. shoulder all these developments and the research?
- development of the SV2, the successor with 800 GFlop/s per chassis, 2nd half of 2002
Surely will the Cray sales force take this chance if and only if the anti-dumping duties vanish. Additionally they have to destroy the bad vector image in the U.S. research institutes that believe in cluster of workstations - my experiences in a Speedup workshop in Switzerland. But there is some need for high-performing vector supercomputers, as can be seen here in Germany. Thus Cray Inc. has a good opportunity but it is in the danger that the company dissipate its strategy and loose a clear focus. Then four different architectures with three (four as UNICOS for the T3E is surely not identical with that of SV1) operating systems and all the applications on these machines have to be supported.
I see an open problem, when the performance of the NEC and the Cray vector come closer with the SV2. Which machine is then the right solution for the customer.
NEC
The Japanese firm made a big deal and now the U.S. market will be open for them. The Cray sales people have good relationships to the big centres and installations and will now offer the long missed high-end vector. This broadens NEC's customer base in the U.S. and as a result of that world-wide. Perhaps it is a chance to enliven the vector architecture.
On the other hand this success is the result of NEC's activity in optimising application software - vectorisation and parallelisation. I often reported on activities in the Swiss Centre in Manno and co-operations for example with McNeal Schwendler, MSC Software package Nastran, where the SX-5 demonstrates extremely good application performance in the Noise, Vibration, Harshness module in the automotive industry. Additionally they port applications in their Stuttgart branch.
The Future
The vector computer market will shrink as it did in the last years. RISC processors are competitive in scalar and some vector problems. The only two, who are committed to build vector supercomputer - NEC through the Japanese Earth Simulator, and now Cray with the SV2, form an alliance to sell high-end NEC's worldwide. Let's see how long these dinosaurs, often predicted as dead, can survive in the future. The only way is to offer better price-performance in specific application areas or program packages as the RISC and Intel competition.
Uwe Harms |
 |