While NEC Corp’s commercial supercomptuers have been uniprocessors, with the SX-3 its first multi-vector machine, NEC is one of the companies in Japan investing heavily in research into parallel technologies. As part of this research it has developed a special-purpose parallel machine called the Cenju named after a Kannon goddess with 1,000 arms. The Cenju consists of 64 processors, eight clusters of eight boards carrying Motorola Inc 68020+68882 and Weitek Corp 1167 chips. Its main application is for hardware logic simulation in developing circuits, especially gate arrays and standard cell design problems that lend themselves to parallel processing – with programs specially written or modified to specify parallelism. Cenju exploits coarse-grained parallelism, the user having specifically to to code what are called parallel synchronisation primitives or functions, which are commands such as Fork/Join, Barrier or Remote Procedure Call. Object-oriented languages hold promise for solving problems on this sort of machine, and a group within NEC software development is working on parallel algorithms based on object oriented languages. The next target of research with Cenju successor machines is speeding them up to several hundred GFLOPS. In other parallel architecture developments, NEC is working as part of the government-sponsored Institute for Computers Of Tomorrow Fifth Generation computer project on massively parallel machines, in particular a Sequential Prolog machine, the CHI-2, which is running a DeoxyrhiboNucleic Acid – DNA – pattern-matching program, involving pattern matches of 3,000m codes. Here the object-oriented stream-based language A’Um is being used. NEC has also developed a specialised architecture for neural network simulation. Another project that stands out as the embodiment of NEC recent Chairman Kobayashi’s dream of C&C – Computers & Communications – is the Mermaid multi-media desktop conferencing system. This enables three or more people, each at different locations, to participate in a tele-conference, viewing each other’s faces in real-time in one window and being able to change documentation and other text with real-time responsiveness so that it appears on the other person’s screen as well.
Recently, NEC invited Anita Byrnes to visit its Central Research Laboratories on the outskirts of Tokyo, where research into key Computer & Communications technologies is done. Here is her report.
Each of the participants’ screens have shared windows for exchanging information, which may be pulled up from commercial databases or other fields such as program libraries; blackboards and notebooks into which entry is made using a conventional keyboard or a tablet and a pen; and dynamic joining and leaving functions. One meeting coordinator is responsible for bringing people toegether and controlling the meeting, but he can pass control to another. The system is said to be more convenient and less costly than teleconferencing which involves special rooms and equipment. The demonstration system was performed on an EWS 4800/220 MIPS Computer Systems Inc R-series RISC-based workstation with scanner and video camera attached, but NEC has already developed a personal computer version that runs on the popular (in Japan) NEC 9801 series. At the moment the system requires a leased line but ISDN should facilitate making it a practical system. The personal computer-based demonstration system is scheduled to be on display at Telecom 91 in Geneva, Switzerland in October this year.
As it fills in background to latest Linpack tests, says others are unfair
Commenting on those benchmark tests at theUniversity of Tennessee by Dr Jack Dongarra (CI No 1,693), NEC Corp says that in both the as is version of the test using a matrix of order 100, and the towards peak performance test with a matrix of 1,000, the NEC SX-3/14 came out on top, with results of 314 and 4,231 MFLOPS respectively, compared with the Cray Research Inc Y-MP/8 with 275 and 2,144 MFLOPS respectively. The Fujitsu Ltd VP2600 scored 249 MFLOPS and 4,009 MFLOPS r
espectively in the tests, the latter being previously the fastest achieved on this test. The performance difference in the NEC compared with the Cray was achieved mainly through improvements to the Fortran compiler and short vector performance. Hitachi’s S-820/80 ran a poor fourth, with matrix-order 100 value of 107 and a logical maximum performance of 3,000. NEC’s criticisms of the original Los Alamos report included the fact that the test had not compared like with like, in that the SX-3 was being compared against a new protoype – NEC is equally confident of its new prototype currently under develoment; and that the previous tests had been developed on a Cray and thus were biased against any other supercomputer.
