Control Data Corporation
|
|
Control Data Corporation, or CDC, was one of the pioneering supercomputer firms. For most of the 1960s they built the fastest computers in the world by far, only losing that crown in the 1970s to what was effectively a spinoff. CDC was one of the eight major computer companies through most of the 1960s; along with CDC these were IBM, Burroughs, NCR, General Electric, Honeywell, RCA, and UNIVAC. CDC was well known and highly regarded throughout the industry at one time, but today is largely forgotten.
| Contents |
Background, origins: WWII–1957
During World War II the US Navy had built up a team of engineers to build codebreaking machinery for both Japanese and German mechanical ciphers. A number of these were produced by a team dedicated to the task working in the Washington, DC area. With the post-war wind-down of military spending the Navy grew increasingly worried that the team would break up and scatter into various companies, and started looking for ways to covertly keep the team together.
Eventually they found their solution; the owner of a Chase Aircraft affiliate in St. Paul, MN was about to lose all his contracts with the end of the war, and after several meetings with increasingly high-ranking Naval officers he eventually agreed to house the team in his glider factory. The result was Engineering Research Associates (ERA), a contract engineering company that worked on a number of seemingly unrelated projects in the early 1950s.
One of these was one of the first commercial stored program computers, the 48-bit ERA 1103. The machine was built for the Navy, who intended to use it in their "above board" code-breaking centers. In the early 1950s a minor political debate broke out in Congress about the Navy essentially "owning" ERA, and the ensuing debates and legal wrangling left the company drained of both capital and spirit. In 1952 the owner sold ERA to Remington Rand.
Although Rand kept the ERA team together and developing new products, they were most interested in ERA's drum memory systems. Rand soon merged with Sperry Corporation to become Sperry Rand, and in the process of merging the companies, the ERA division was folded into Sperry's UNIVAC division. As the Sperry "big company" mentality encroached on the decision-making power of the ERA founders, they eventually got fed up and decamped to form Control Data in 1957, setting up shop in an old warehouse down the road in Minneapolis. Of the members forming CDC, William Norris was the unanimous choice to become CEO of the new company. Seymour Cray was likewise chosen to be the chief designer, but was still in the process of completing the 1103-based Navy Tactical Data System (NTDS) and did not leave to join CDC until this was complete.
Early designs; and Cray's big plan
CDC started business by selling parts, mostly drum systems, to other companies. Cray joined the next year and immediately built a small transistor-based 6-bit machine known as "Little Character" to test his ideas on large-system design and transistor-based machines. Little Character was a success, and they soon released a transistorized version of their 1103 design as the CDC 1604 in 1959, with the first machine delivered to the US Navy in 1960. A 12-bit cut-down version was also released as the CDC 160A in 1960, arguably the first minicomputer. New versions of the basic 1604 architecture were re-built into the CDC 3000 series, which sold through the early and mid-1960s.
Cray immediately turned to the design of a machine that would be the fastest (or in the terminology of the day, largest) machine in the world, setting the goal at 50 times the speed of the 1604. This required radical changes in design, and as the project "dragged on" (it took about four years in total) management became increasingly upset and demanded ever-greater oversight. Cray eventually got fed up and 1962 demanded to be set up in his own remote lab, or he would quit. Norris agreed, and after a short search the team moved to Cray's home town of Chippewa Falls, WI.
The CDC 6600: defining supercomputing
In the new lab Cray, Jim Thornton and Dean Roush put together a team of 34 engineers (themselves included), and continued work on the new design. In 1964 this was released as the CDC 6600, outperforming everything on the market by roughly ten times. The 6600 had a simple CPU, but used a series of external I/O processors to offload many common tasks. That way the CPU could devote all of its time and circuitry to processing data while the other controllers dealt with mundane tasks like punching cards and running disks. Using late-model compilers the machine racks up .5 MFlops, while hand-coded assembly can deliver about 1 MFLOPS, impressive considering that it is about 40 years old. A slower version was released as the CDC 6400; a two processor version of that was called the 6500.
It was after the delivery of the 6600 that IBM took notice of the new company. At the time Thomas J. Watson, Jr. asked (paraphrased) how is it that this tiny company of 20 people can be beating us when we have thousands of people?, to which Cray quipped you just answered your own question. In 1965 IBM started an effort to build their own machine that would be even faster than the 6600, the ACS. Two hundred people were assembled on the west coast to work on the project away from corporate prodding, in an attempt to mirror Cray's offsite lab. The project produced interesting architecture and technology, but was not compatible with IBM's very successful System/360 line. It was redirected to be System/360-compatible, but this compromised performance, and ACS was eventually cancelled in 1969 after producing no product.
In the short term IBM also went ahead and announced a new version of the famed System/360 that would be just as fast as the 6600, the Model 92. This machine didn't exist, but that didn't stop sales of the 6600 drying up while people waited for its release (a tactic known today as FUD). Norris didn't take this lying down, and a year later filed an anti-trust suit against IBM, eventually winning over 600 million dollars and picking up the Service Bureau Corporation, an IBM subsidiary that ran computer processing for other corporations on SBC's computers.
The CDC 7600 and 8600
The same month they won against IBM they also announced their new machine, the CDC 7600. Cray had started the design even before the 6600 was shipping, and allowed it to mature fully. This machine ran at about ten times the speed of the 6600. Much of this speed increase was due to extensive use of pipelining, a technique that allows different parts of the CPU to work on different parts of the instruction processs at the same time. The time to run any particular instruction is no faster, but the program as a whole moves through the system more quickly as the instructions are cued up. Sadly the complexity also led to poor reliability, and while the 7600 project certainly paid for itself, it did serious damage to CDC's reputation.
Cray then turned to the latest in the series, the CDC 8600. The 8600 was essentially four 7600's in a single, much smaller, case. The smaller size and shorter signal paths allowed the 8600 to run at much higher clock speeds, which was combined with higher speed memory for most of the performance gains. Unfortunately the 8600 was "old school" in terms of physical construction, using individual components soldered to cards. There were so many solder joints on the 8600 that the machines never worked reliably, one bad joint and the machine was "flaky". Cray decided that a re-design was needed.
The STAR and the Cyber
In addition to the redesign of the 8600, CDC had another project called STAR underway, led by Cray's former collaborator on the 6600/7600, Jim Thornton. Unlike the 8600's "put four 7600s in a box" solution to the speed problem, the STAR was a new design using a technique we know today as a vector processor. By highly pipelining math instructions with purpose-built instructions and hardware, math processing could be dramatically improved in a machine that was otherwise slower than a 7600. Although the particular set of problems it would be best at was limited in comparison to the "generalist" 7600, it was exactly these sorts of problems that customers bought CDC machines to solve.
The two projects competed for limited funds in the late 1960s, and Norris felt that the company could not support development of the STAR and a complete redesign of the 8600 at the same time. He told Cray, who in 1972 Cray left to form Cray Research. Norris remained a staunch supporter of Cray, and even invested money into his new company. Eventually the 8600 was cancelled in 1974, and the STAR released the same year. The STAR proved to have considerably worse "real world" performance than expected, eventually leading to Jim Thornton's dismissal.
A variety of systems based on the basic 6600/7600 architecture were re-packaged at different price/performance points as the CDC Cyber, and became CDC's main product line in the 1970s. This included an updated version of the STAR architecture as well, the Cyber 205, which had considerably better performance than the original. But by this time Cray's own designs like the Cray-1 were using the same design techniques as the STAR, yet doing it much faster due to various design details. Sales of the STAR and its follow-ons dropped, and CDC found itself being pushed out the supercomputer market by the late 1970s.
ETA Systems, hard disks, oblivion
CDC decided to fight back, but Norris agreed with Cray in thinking that the company had become too ossified to be able to quickly design anything competitive. Instead he set up a new spinoff company in 1983, ETA Systems, their design goal being a machine able to process at 10 GFLOPs, about 40 times the speed of the Cray-1. The ETA design never fully matured and was unable to hit its design goals, but was nevertheless one of the fastest computers on the market and a handful were sold over the next few years. Eventually the effort ended after half-hearted attempts at selling ETA, and in 1989 most of the employees were laid off and the remains of the company were re-folded into CDC.
Meanwhile several very large Japanese manufacturing firms were entering the market as well. The supercomputer market was too small to be able to afford more than a handful of players, and CDC started looking for other markets. One of these was the high-performance hard disk market, which was becoming more interesting as personal computers started to include them in the mid-1980s. By the early 1990s CDC was a major player in the hard drive market, their Wren series drives were particularly popular for "high end" users. They also co-developed the now-universal ATA interface with Compaq and Western Digital to lower the cost of adding low-performance drives.
Oddly, then, CDC decided to exit the hard drive business entirely in 1992, and sold off their line to Seagate, who had been seriously lagging in the high-end drive market. The remains of CDC became known as Control Data Systems, Inc., now Syntegra (USA), a subsidiary of the BT Group's BT Global Services; CDC's non-computer business became the Ceridian Corporation.
External links
- History of Large Computers at LLNL (http://www.nersc.gov/~deboni/Computer.history/Page4.dir/images/comp.history.chart.jpg) (JPEG) – From George A. Michael et al's compilation "Stories of the Development of Large Scale Scientific Computing at Lawrence Livermore National Laboratory" (LLNL's CDC computers' performance are shown by red bars on the graph)
- Control Data Corporation Records (CBI 80) (http://www.cbi.umn.edu/collections/inv/cdc/cbi00080.html) – At the Charles Babbage Institute, University of Minnesota, Minneapolis; CDC archives collection donated by Ceridian Corporation in 1991 (pages also contain historical timeline, product timeline, acquisitions list, joint ventures list)
- Yahoo! - Ceridian Corporation Company Profile (http://biz.yahoo.com/ic/10/10402.html)
- Yahoo! - Syntegra (USA) Inc. Company Profile (http://biz.yahoo.com/ic/11/11174.html)
