Seymour Cray, Super Computist

b 1925, d 1996 Colorado Springs

Seymour Cray died of injuries received in a Sept 22, 1996 car crash in Colorado Springs. One car had made a sudden lane change, forcing another car into Cray's. His Jeep Cherokee went off the road and rolled three times. He suffered serious, and ultimately fatal, head injuries. He survived for two weeks in the hospital but never recovered consciousness. He had been putting together a new company, SRC Computer, after the failure of Cray Computer and the Cray-4.

What a loss! Cray was one of the best of our field, and a giant of post-war electrical engineering as a whole. He started in the early transistor era in the 50s, designed the first big vector machines, the CDC 6600 and 7600, while at Control Data, then spun off Cray Research to do the Cray-1 and Cray-2. His machines were the fastest computers in the world from the mid-60s to the mid-80s, when Steve Chen's Cray-XMP superseded his own design. In these days when a hundred thousand people are involved in computer engineering, no other person is ever likely to dominate the field that way again.

A few bits of Cray lore:

• Of all his many patents, he was proudest of one he received for a magnetic amplifier design. This was a means of amplifying signals without tubes or transistors, but just by exploiting the non-linear saturation curve of an iron transformer core. When the design came back from the patent office it read "No previous art".
• The Cray-1 wasn't just the fastest vector machine on the planet, it was also one of the most efficient random integer code machines. The vector pipeline was deep, but the integer pipeline was short. It could do character string moves more efficiently on a cycle-by-cycle basis than the VAX microcode could. This realization in the early 80s was when the VAX architects knew they were doomed, and was one of the spurs to the RISC revolution.
• For relaxation he liked to build sailboats. He didn't like to sail them; that wasn't the point. In fact, at the end of the season he would burn his old boat to make room for the new.
• When people asked why he didn't use caches, he replied "You can't fake memory bandwidth that isn't there." No cache could handle the data sets that his machines were meant for. Instead, he used hundreds of megabytes of interleaved 20 ns ECL RAM.
• He had the same cavalier attitude towards floating point precision; the bottom few bits of his double precision numbers were not at all guaranteed to be right. For instance, he did divides by a Newton-Raphson iteration instead of the bit-at-a-time method that gives accuracy to the last LSB. If you need those bottom few bits, you can't be sure that even double precision is enough. Use the right algorithm instead of crippling the hardware.

So his machines were not easy to use or program by today's standards. They were thoroughbreds, meant to solve problems of literally national survival, like bomb design, aerodynamics, and cryptography. When the Cold War ended, his last, brilliant works, the Cray-3 and Cray-4, no longer had a market. Sorry as we may be to see them fail, we can't miss the conditions that made them needed.

A few links to him:

A Tribute to Seymour Cray by a colleague, Charles Breckenridge, at his last company, SRC Computers.

Cray Interview by the National Museum of American History

A Seymour Cray Perspective by Gordon Bell