NEC Corp’s Supercomputer Systems division is ploughing ahead in its quest to bring the supercomputer down to departmental level with the launch of the increased memory SX-4A – but the US might miss out. NEC says by using CMOS technology and less expensive memory, the SX-4 has now become a viable proposition at departmental level, […]
NEC Corp’s Supercomputer Systems division is ploughing ahead in its quest to bring the supercomputer down to departmental level with the launch of the increased memory SX-4A – but the US might miss out. NEC says by using CMOS technology and less expensive memory, the SX-4 has now become a viable proposition at departmental level, and is a serious commercial tool rather than a toy for the scientific world. The SX-4A is the latest addition to the SX-4 family of shared memory vector supercomputers announced two years ago (CI No 2,539), and provides four times the memory per CPU of the current models, the company says. This means it will deliver a maximum of 32Gb of memory. The company has had a big sales push into Europe this year, and so far has sold 12 machines into 10 sites. It reckons it has seven SX-4s in the US to date, and 63 in Japan. It is based mainly on CMOS technology, and NEC says the cost of ownership of an SX-4 is about 40% of that of a similar Cray Research Inc supercomputer. An entry-level SX-4A with one or two CPUs, giving a capability of up to 3.6 GFLOPS, will cost around $500,000. NEC says until now, the sort of memory demands on a supercomputer could be met by only costly synchronous static RAM. In fact the demand placed on memory by the combined performance of up to 32 SX-4 CPUs performing at 64 GFLOPS in a shared memory node was compared to the task of simultaneously hammering a nail in on every one of the 15,800m inches of the equator, every second. The SX-4A uses synchronous dynamic RAM, which it says will satisfy the growing demand from customers for much larger, affordable memory.
By Joanne Wallen.
Dynamic RAM is not only cheaper than static RAM, but also takes up less space, since it generates less heat and needs less cooling. While it may be slightly slower at the very high end, in the main this is compensated for by the ability to add more memory per CPU. Customers therefore have the choice between the SX-4 and 4A, depending on the memory requirements of their applications. European customers include car manufacturers Daimler-Benz AG and Volkswagen AG, the Dutch and German aerospace research centres NLR and DLR, and the Danish meteorological office. Volkswagen has three departmental SX-4s, on which it designs different parts of the car. Dr Keith Corless, national sales manager for NEC European Supercomputer Systems says the beauty of the shared memory architecture is the availability of application software. Code will run on the SX-4 without the sort of parallelisation changes needed to run on massively parallel systems. However, the SX-4 will also support parallel code, so users could parallelize their code incrementally. He believes that one of the main reasons there have been initiatives such as the European Union’s Europort, to get more parallel code (CI No 2,876), is that when these were launched, no one would have thought there’d be TeraFLOPS processors in 1995. The company says that since the vast majority of parallelized industrial applications such as structural engineering can make efficient use of only between eight and 32 CPUs, distributed architectures like massively parallel processing cannot provide access to large memories without substantial performance loss caused by remote memory access. So why might the US miss out? Unfortunately, as a result of the anti-dumping complaint launched by Cray this year, whereby Cray claims NEC sold four supercomputers for the price of one when bidding for the US National Center for Atmospheric Research (CI No 2,983). It looks as if all orders are on hold until the issue is resolved. NEC claims it has not had a chance to put its case yet. Its argument is that Cray’s allegations were based on its own estimates of the costs of a supercomputer. NEC says since it shares technology across its company, and development for one technology is deployed elsewhere in the group as well, its costs are far lower than those of Cray. Also, the cost savings of using CMOS technology, especially when it uses its own NEC CMOS technology, are pro
bably not fully appreciated by its competitors, it said. Both NEC and its potential US customers will have to wait for the outcome of the US Commerce Department and International Trade Commission enquiry. The SX4-A ships in March.