Cray-3

from Wikipedia, the free encyclopedia
Cray-3 "Brick" processor

The Cray-3 supercomputer is Cray's successor to the Cray-2 . The system should be the first use of gallium arsenide - semiconductors (also called GaAs semiconductor) in the electronic data processing to be what had already been tried in the previous model. The project turned out to be a huge flop, however, as the Cray-3 was very expensive due to its gallium arsenide circuitry and therefore only one copy was shipped.

history

As soon as the Cray-2 was completed, Seymour Cray began developing the Cray-3. With this computer he left Cray Research and founded his third company, Cray Computer Corporation (CCC for short). It was a peaceful breakup. The reason: Due to the two-year delay of the Cray-2 , which was to appear in 1982, it was necessary to establish a successor to the Cray-1 - after all, the competition is also bringing out new models. The choice fell on a further development of the Cray-1 to 2 processors, the Cray X-MP . It was possible to develop this further (later the Cray Y-MP with 4 or 8 processors and the Cray T90 with 16 processors emerged from it), but no company can practically afford to run two supercomputer lines: Such computers are sold individually and rarely more than 100 pieces of one model.

The problem that had already delayed the development of the Cray-2 was the cycle times: Slowly one came to the limits set by physics. The light and thus also the clock signals move only 30 cm in a nanosecond. Transistors switch in ranges that are still below 1 ns, but when data are processed, the data pass through a few dozen gates, so that the switching times add up.

Seymour Cray saw the solution in another semiconductor material: gallium arsenide . Silicon is the most commonly used material and also the cheapest. But it is the worst of the electrical properties. Both germanium transistors and semiconductors made from elements of the third and fifth main groups such as gallium arsenide switch considerably faster because the electrons have a higher mobility. For reasons of price and practicality, germanium is out of the question for highly integrated circuits, gallium arsenide remained.

What Seymour Cray underestimated was the complexity of the task: Developing a new computer is difficult. But it was too difficult to construct new circuits on the basis of a material that was not generally used. The development was delayed from 1988/1989 to 1993. Secured by an order and state support, he was able to present a processor of the Cray-3 in 1993 at a cost of 120 million USD and also install it in the NCAR . The entire system was to consist of 16 processors in a 90 × 90 cm cube, 4 GB of memory and direct connections with 8 GB / s between the processors. With a cycle time of 2.11 ns, one processor achieved a peak performance of 0.948 GFLOP, the entire computer 15.17 GFLOP. The power consumption of the small cube was 88 kW.

But after the Cray-3 was ready for production, nobody wanted it - it was simply too expensive. The same thing happened with the CM-2 from Convex, another supercomputer manufacturer that used gallium arsenide.

Cray Research had further developed the architecture of the Cray-1 and with the Cray C90 a 16-processor system with a performance of 1 GFLOP per processor - with a cycle time of just 4.2 ns and conventional technology. In 1995, Cray's CCC went bankrupt. She couldn't sell a Cray-3 system. In the meantime, however, Seymour Cray was already working on the Cray-4 - also with gallium arsenide and 1 ns access time. On September 22, 1996, he died of a traffic accident before he could complete the Cray-4.

System information

Parameter Data
Machine type Shared-memory multi-vector processor
operating system UNICOS (Unix variant from Cray Computer Corporation)
Compiler Fortran, C.
Cycle time 2.11 ns
Theoretical peak performance:  
Per processor 0.948 Gflop / s
Maximum (16 processors) 15.17 Gflop / s
Main memory 4 GB
Memory bandwidth:  
Bandwidth of each processor 8 GB / s
Number of processors 1-16