Cray Research

Cray went ahead with the design of the 7600 and the 8600 but CDC decided not to market the 8600 and so Cray became unhappy yet again.

CDC had become a power in the business computing field and they downgraded the work on supercomputers. This time Cray really did have to go it alone. He left CDC and formed Cray Research - dedicated to supercomputers. The leaving wasn't acrimonious, CDC contributed $500,000 to the startup and they promised not to build any more supercomputers.

The design Cray had in mind was revolutionary and would eventually make money - but at first it was a struggle. Cray Research was set up in 1972. By 1976 the money had run out and the Cray 1 still wasn't ready.

The solution to the problem was to go public but Cray Research was a company that didn't have a product and it did have a $2.5 million deficit. Amazingly Wall Street came up with the funding - $10 million almost immediately - the word about Seymour Cray must have leaked out.

Cray would work with his team for a few hours and then go home at around 4pm - but he would return at night to work alone at the late shift. An eccentric genius who surely must be worth a few dollars gamble.

Photograph courtesy of the Charles Babbage Institute, University of Minnesota, Minneapolis

The first Cray-1 was delivered to Los Alamos Lab in March 1976 at a cost of $8.8 million. There were thought to be only eight potential customers for the Cray-1. By 1980 it had sold nine, thirteen more in 1981, and fifteen more in 1982. Not a mass market product but it made the company a great deal of cash which more than repaid the investment.

Speed Counts

The reason that people wanted the machines was simple - it was the fastest machine available. By using a vector processor it could perform the sort of arithmetic needed to solve many scientific problems.A whole set of numbers could be multiplied in a single operation.

As well as being a fast vector processor the Cray-1 was also the fastest scalar processor of its time. Its cycle time of 12.5ns was equivalent to a clock rate of 80MHz. The only real competitor to the Cray-1 was the CDC Cyber 205 which worked at only 40MHz. Yes, that's right, CDC did not keep its promise to stay out of the supercomputer market.

Although these clock rates are not high by today's standards you have to take into account the fact that the machines were built of individual bipolar logic chips. This results in parts of the machine that would be next to each other on a single chip using today's large scale integration were separated by large distances on printed circuit boards. In this case the transmission time for a pulse is an important factor in how fast a machine can work. Pulses cannot travel faster than light - around 1 foot per nanosecond - and in practice they usually propagate at closer to .3 to .9 feet per nanosecond.

The Cray 1 achieved its higher clock rate by being smaller and packing 300,000 chips into less than 100 cubic feet. To get rid of the heat the whole machine had to be cooled by a Freon refrigerant - something Cray pioneered in his days with CDC. Each circuit board was bonded to a sheet of copper and which slid into massive vertical slabs of aluminium in which the Freon circulated.

The machine was also built into a semi-circular case with what looked like seats around the base. The horseshoe shape kept the connections from one part of the machine to another short and the "seats" were in fact the power supplies. Even if the design was functional there is no denying that it looked good and it is still instantly recognisable as the archetypal supercomputer.

The difficulties in building the machine must have been huge. With over 300,000 wire interconnections, each trimmed to an exact length so as to produce a fixed travel time for the pulses, which ran between 1,600 circuit boards, it must have been a wireman's nightmare.

Although it is interesting to know what the clock rates were, it is important to realise that the actual speed of computation had more to do with the overall design of the machine. As well as a single high speed processor, the Cray 1 also had a high speed vector processor which could perform more than one operation at a time. Although it is difficult to be precise about the power of the Cray-1. claiming a top speed of 100Mflops (i.e. 100 million floating point operations per second) is reasonable.

Number 2 and beyond

With the success of the Cray-1 the next model had to be faster still. The Cray-2 was to be smaller and this needed new a hardware design. Using special 3D modules to hold the chips Cray managed to pack 250,000 devices into a horseshoe shaped cabinet even smaller than the Cray-1. Every connection was reduced to less than 16 inches and the result was a cycle time of 4ns i.e. a clock rate of around 120MHz. Half way through the project Cray decided to switch from bi-polar chips to cooler CMOS devices. Cooler they may have been but getting rid of the heat was still a problem. This time he decided to flood the machine with Fluorinert - which also resulted in the Cray-2's nickname "the computer in the aquarium".

Cray Research was selling more and more supercomputers. Partly because falling component costs allowed it to cut the price of the machines by $5 to $10 million. Seymour Cray was once again involved in an important successful company and he wasn't happy. It just didn't leave enough time for building supercomputers.

In 1981 he left the board of Cray Research and went to work again at Chippewa Falls. He had an agreement with the company that he could pursue any line of development even if Cray Research wasn't interested in it. It looks as if he had found paradise for the second time. But Cray Research needed to carry on making money so in 1985 the Cray X-MP was launched. Its design was mostly the result of the work of Steve Chen and not Seymour Cray who was busy at work on the Cray-3.

The Cray-3

The Cray-3 was eventually delivered in 1993 -  a small, four foot high grey box with none of the charisma of the earlier machines. It was revolutionary in that it used gallium arsenide semiconductors in its chips for speed and low power. Even though its 128M words of memory, four processors and 20Gbytes of disk storage may not seem impressive today the 90Kwatts of power it consumed is. The Cray 3 wasn't a great sucess and it probably was the cause of the eventual collapse of the company in 1995. The machine was good but times had changed - the easing in the cold war resulted in a reduced need for supercomputers and the age of the massively parallel computer was underway.

Seymour Cray was perhaps the last of the lone supercomputer designers. His designs depended on size and engineering smartness. Today's supercomputers depend on single chip processors and use large interconnected arrays to create parallel machines. Designing one doesn't need the sort of skills that Cray had in plenty. It isn't work that a single man could or would want to carry out and group work was never his strong point.

In the afternoons he would disconnect his phone. He would even refuse invitations to speak at scientific gatherings. He believed that a supercomputer was best built by one man working alone and keeping the whole design inside his head. This no doubt resulted in his reputation as a recluse and an eccentric - but you can only envy his freedom. When taking his annual holiday to the Tropics he would buy a new car. Rather than be side tracked into the time wasting chore of choosing one, he simply picked the one nearest the right of the showroom door.

<ASIN:0387982698>

<ASIN:0412598108>