'Gigaflops battle’ resumes

Press, 16 May 1989, Page 27

'Gigaflops battle’ resumes

The “battle of the gigaflops” is under way again. The five leading makers of supercomputers — two American firms and three Japanese — are outshouting one another with mostly meaningless claims. Their boasts are about whose machine holds the record for data processing. Recently a flurry of announcements came from Japan as Fujitsu, Hitachi and NEC jostled for position at home while preparing for the next move by the most successful firm in the business, Cray Research of Minneapolis. In the early 1980 s supercomputers — former military megamachines that were once used to crack codes and simulate nuclear explosions — came in from the cold. Today they help engineers to design better cars, chemists to probe molecular structures and meteorologists to forecast (patchily) the weather months ahead. It takes a supercomputer minutes or hours to do what its smaller brethren would — at best — take days to accomplish. The market for these Goliaths of computing (costing SSM to S2OM) is growing at nearly 40 per cent a year. Last month NEC broke ranks with its Japanese rivals and entered the fray with the country’s first true “multiprocessor” design. Its new SX-3 links four central processors to give, so the company claims, a peak processing speed of 22 gigaflops. This means that in one second it can do 22 billion float-ing-point calculations — the sort of sums by which computer power is measured. In terms of gigaflops, NEC’s new machine is six times faster than Cray’s Y-MP, and twice as fast as the former record-holder, the ETA-10 built by ETA Systems, also American. The salvo of claims and counterclaims is more a “brochure battle” than a real clash of metal on metal. The new NEC machine will not be on the market until mid-1990, by which time a 16-pro-cessor Cray-3 is supposed to have made its debut. Besides, peak processing power reveals little about a supercomputer’s usefulness, or even its over-all speed. Much depends on how it is configured and what software it uses. Good architecure makes a difference. It is better to have several central processors working in parallel as a team, than to try and wring performance from a single processor slogging through a problem one step at a time. Designing a multiprocessor supercomputer is tricky enough. Writing its software can be mindnumbing. For one thing, a different algebra is needed — one that can manipulate whole columns of data (i.e. vectors) at once, instead of simple numbers (i.e. scalars) one at a time. For another thing, real problems rarely take kindly to vector processing. Few can be expressed wholly in the language of vectors, and thus be crunched by parallel processors working flat out; much remains to be processed messily in a conventional step-by-step manner. The trouble with boasting about gigaflops is that such measures refer only to the part of a calculation that can be

expressed in vectors. So far American designers of supercomputers have had an arm-lock on parallel processing and on the complex programming needed to make it work well. Cray’s Y-MP and the ETA-10 both use eight parallel processors. Such designs, allied to internal software that generates (or “compiles”) the coded instructions for vector processing, are what give a supercomputer its appetite for data. Thanks to its able compiler, even the old Cray-1, running at a mere 160 megaflops (million floating-point operations per second), can chew its way through problems faster than rival machines with speeds of over 400 megaflops. Although they are oldfashioned in design, Japanese supercomputers have made up for their lack of architectural finesse with horsepower. Fujitsu has excelled at designing high-speed logic circuits based on gallium arsenide (trickier to use but faster than silicon) as well as the even more exotic “high-electron-mobility” transistors (which use extra-ener-getic electrons) that few other firms understand.

Hitachi, which until recently carried the flag for Japanese supercomputers with its 3 gigaflop S-820, got its raw speed from a single central processor stuffed with extremely fast memory circuits. Four months ago Fujitsu announced a 4-gigaflop design called the VP-2000 which it will introduce next year. Meanwhile, ETA Systems holds the speed record with its 10gigaflop machine. Remarkably, all Japanese supercomputers have used a single processor. To catch up with the Americans in parallel processing, the Japanese Government is sponsoring research that aims to have a 10-gigaflop multiprocessor machine by early 1990. Fujitsu has been developing the hardware under contract to the Ministry of International Trade and Industry. Yet the Ministry’s flagship computer project has been pipped by a Japanese firm that has taken its own road to parallel processing. The announcement by NEC is significant not just because its machine is Japan’s first commercial attempt to build a multiprocessor computer for

parallel processing. More important for potential sales is the fact that the SX-3’s operating system (the software used for managing a computer’s internal operations) is based on Unix. To steal some of NEC’s thunder, both Fujitsu and Hitachi announced a few days earlier that they would equip their own supercomputers with Unix. Previously Japanese supercomputer. makers stuck to proprietary operating systems, hoping thereby to addict their customers to pricey software written in Fortran, the programming language favoured by physicists and mathematicians. They have even produced compilers to allow customers’ programs written in Fortran to be speedily translated into the language of vectors. This has had some success in Japan. Elsewhere, the trend has been to program supercomputers in. While NEC has begun to follow the parallel-pro-cessing route pioneered by Cray, American firms have started borrowing ideas from Japan. The Cray-3 will use microchips made of gallium arsenide instead of silicon, packed

in far denser arrays and made to flip on and off in two nanoseconds (billionths of a second) — ten times faster than their predecessors. What bothers both Cray and NEC more than the thought of technical stunts that the other might pull is what IBM is secretly up to. With many irons of its own in the supercomputer fire, the world’s largest computer company has become an active investor in Mr Steven Chen’s Supercomputer Systems Inc. Mr Chan fell out with Cray two years ago and took 45 engineers with him to develop a parallel computer with up to 256 processors and speeds of up to 128 gigaflops. Behind Mr Chen are some three dozen young American companies with new ideas for supercomputers, including machines with up to 65,000 parallel processors, the so-called “finegrained” approach. With so many developments to come, the speed records for supercomputing will soon change hands so fast that techno-nationalism itself will become a blur.

—Copyright, the “Econo-