What ever happened to…
IBM’s STRETCH Supercomputer
by John C. Dvorak
When few others at IBM thought it was important, Stephen W. (Red) Dunwell conceived of a computer that would be a “giant step” forward not only for IBM, but for computer design in general. And with a combination of vision, chutzpah, and patience, he led a project that designed and built the world’s fastest and most technically advanced computer: the Stretch supercomputer. But within a month of delivering the prototype, an apparently embarrassed Thomas J. Watson, Jr., publicly declared the project a failure. In fact, it wasn’t.
In late 1954, Dunwell and Werner Buchholz began writing a series of memos about a conjectural computer system they dubbed “Datatron” that would supplant the soon to be completed Type 704 and 705 computers. Most of the memos dealt with technical issues, but a memo labeled number zero served as something of a manifesto. It advocated a computer which would “assure IBM a pre-eminent position in the field of electronic data processing” by taking a “giant step and making substantial advances on all fronts.”
The memos led to occasional meetings which were attended at various times by, among others, John von Neumann, who was a consultant to IBM at the time, Gene Amdahl, who led the design work on the IBM 704 and, later, the System/360; and John W. Backus, who was the head of the FORTRAN team.
Dunwell was careful to tune his message to what corporate ears wanted to hear and so, when word got around that Livermore Radiation Lab (LRL), in California, was very interested in obtaining a high-performance computer, he began to slant his memos toward performance issues.
Once he claimed that the Stretch’s performance would be 100 to 200 times that of the 704. Those claims eventually came back to haunt him, but the boasts got him the attention he needed within the executive ranks to get himself the project leader position over his peer Gene Amdahl.
At the time IBM was falling behind in scientific computers, so a meeting was convened in January 1955, to consider the next large technical computer that IBM would produce. “Datatron” was among the alternatives being considered.
By the end of the meeting, it had been decided that Engineering would pursue a version of the “Datatron” that would include a fast arithmetic unit and transistorized registers that would allow for a small, fast memory to reduce the number accesses to the larger, slower main memory. The main memory itself would be larger to reduce the number of even slower input/output operations.
To find government funding to develop a prototype the project leaders first approached Edward Teller at Livermore Radiation Lab (LRL). It was known that LRL was in the market for a fast computer. But IBM wasn’t able to fulfill the contract within the time that LRL needed such a machine and the contract eventually went to Remington Rand. The corporate wheels, however, had been set in motion.
During the early development period Amdahl took a prominent role in outlining system characteristics. However, this period was mainly aimed at selling IBM’s top management on the project. Management, however, wasn’t sold until a contract had been signed with Los Alamos Scientific Laboratory (LASL) for what would now be known as Project Stretch. (The name came from the project’s attempt to “stretch” the current limits of computer design.)
With funds secured, the real challenge was now to find a way to fulfill the promises. Although the specifications to which Los Alamos had agreed didn’t mention “throughput,” a cover letter for the proposal promised “a speed at least 100 times greater than that of existing machines.” In attempting to accomplish this goal, Dunwell’s team had to overcome numerous engineering problems including designing and fabricating a load-sharing switch which would allow them to use transistors to drive the large ferrite-core memory.
Unfortunately, as all the engineering problems were being resolved, the Stretch team’s own hype was about to catch up with it. The originally scheduled delivery date for the Stretch supercomputer was May 1960. But in December 1959, an internal technical audit foresaw a delay and suggested that announcement of the 7030, a planned line of Stretch-like computers, be delayed until the Stretch had been tested and delivered to Los Alamos.
But fears that any further delay would hurt IBM’s position in the scientific market and kill any chances of recouping the now huge development costs of the project forced Thomas J. Watson, Jr., to pre-announce the 7030 at the annual stockholder’s meeting
in April 1960.
Although expectations were now that the Stretch would operate at 75 times the speed of the 704, even this proved too optimistic. Testing in February 1961, showed that although the Stretch’s internal operations, such as “add,” multiply,” and “subtract,” were mostly in line with Watson’s more moderate expectations, the actual execution of user programs was only about half the expected speed: 32 to 40 times the speed of the 704. Watson, who in the last months of the project was getting regular updates, and was very upset by what he perceived to be deception.
In Dunwell’s defense and in spite of development problems, the Stretch was the fastest machine yet built, an accolade it would retain until Seymour Cray built the Control Data 6600 several years later. And its users were quite happy with its performance, accuracy and reliability.
In May 1961, a month after the Stretch was delivered to Los Alamos, Watson announced at the Western Joint Computer Conference that the 7030 did not perform up to expectations and that, therefore, the price had been reduced from $13.5 to $7.78 million and would only be offered to the eight potential customers with whom negotiations were currently underway. (All of them eventually bought a 7030 and it’s apparent that the company could have sold more.)
Dunwell’s overly optimistic predictions made him a natural scapegoat for the delay’s, cost overruns and less-than-expected performance. He was soon banished to a research lab until hindsight made the Stretch’s accomplishments clear. Dunwell was eventually made an IBM Fellow for his contributions to the Stretch project. At an awards banquet, in April 1966, Watson (who was prone at times to what has been termed “public confession”) apologized for being unfair to Dunwell. Even so, Dunwell soon took early retirement.
Although IBM was estimated to have lost $20 million on the eight 7030 machines that were eventually built, IBM gained a number of benefits from the project. Among the benefits was the hiring of talented young engineers who were attracted by the opportunity to work on “sexy” new equipment. Furthermore, the Stretch’s most successful features were later incorporated into the hugely successful System/360. Those advances included pre-fetch memory, memory protection and program-controlled interruption that allowed for multi-tasking, hardware-based error-detection and correction, and an expanded instruction set.
Many of the engineering concepts found in the Dunwell machine are just now finding their way into today’s microprocessors.