CDC 8600
| Industry | Computer hardware, Supercomputers |
|---|---|
Key people | Seymour Cray (project leader for CDC 8600) |
| Products | CDC 6600, CDC 7600, CDC 8600, CDC STAR-100 |
The CDC 8600 represented Seymour Cray's final supercomputer project at Control Data Corporation. Designed as the follow-up to the CDC 6600 and CDC 7600, it targeted a performance approximately ten times greater than the 7600, which was the leading computer speed at that time. The architecture integrated four 7600 processors within a single compact enclosure to achieve faster clock rates.
Initiated in 1968, development soon encountered significant obstacles. The tightly packed configuration led to problems with reliability and heat dissipation. By 1971, financial difficulties at CDC combined with persistent design challenges prompted Cray’s departure in 1972. Ultimately, the initiative was terminated in 1974 as CDC redirected its efforts toward the CDC STAR-100 lineup.
In the early 1980s, Cray revisited this approach with the creation of the Cray-2, utilizing advances such as integrated circuits and liquid cooling to overcome prior limitations. The Cray-2 bore strong physical and conceptual similarities to the 8600.
Design
During the 1960s, computers were constructed from electronic components mounted on circuit boards assembled into modules. To achieve high speeds, modules needed to be small to minimize signal travel distance; however, increasing module complexity risked degrading performance due to longer internal wiring.
Cray’s solution was to produce larger, densely packed modules while reducing overall machine size. Each module contained eight four-layer circuit boards measuring approximately 8 by 6 inches, stacked similarly to a thick book and consuming roughly three kilowatts of power. These modules were installed within a compact, sixteen-sided cylindrical chassis about one meter in diameter and height, supported by a ring of power supplies—a design feature later echoed in the Cray-2.
Effective cooling was essential given the high power density. Refrigeration specialist Dean Roush engineered copper sheets embedded within circuit boards to conduct heat toward a copper block cooled by a freon refrigeration system. This approach resulted in each module weighing around fifteen pounds, with an external cooling apparatus larger than the computer itself.
The central processor employed Emitter Coupled Logic (ECL), raising the clock frequency from the 7600’s 36.4 MHz to 125 MHz (an 8 ns cycle), representing nearly a fourfold speed increase. Main memory also transitioned to ECL technology and consisted of a substantial standard size of 256k words (2 MB) distributed over 64 banks for rapid access times of approximately 8 ns per word. A high-speed core memory with a slower 20 ns access time functioned as backup storage.
The architecture featured four complete CPUs sharing common main memory. It supported an operational mode that dispatched a single instruction simultaneously to all processors but allowed each to operate on different data (now recognized as SIMD), reducing memory traffic and enhancing throughput. Each CPU ran roughly two and a half times faster than a single 7600, achieving an aggregate processing rate near 100 MFLOPS.
To meet government mandates related to ASCII text processing, the system adopted a 64-bit word format (comprising eight 8-bit characters) instead of the previous 60-bit word length. Instructions were either sixteen or thirty-two bits long, accompanied by a new set of sixteen general-purpose X registers replacing earlier register configurations. The Peripheral Processor subsystem managing input/output remained largely consistent with prior models.
Efforts to maintain compatibility were constrained by changes such as word size modifications; however, floating-point formats were preserved to allow straightforward porting of Fortran applications.
Company challenges
In 1971, Control Data imposed a 10% salary reduction across its divisions as a measure to offset expenses related to an ongoing legal battle with IBM. Cray sought an exemption for his division to finalize the shipment of the 8600, but management denied this request, leading him to voluntarily cut his own pay down to minimum wage.
By 1972, the 8600 had become so unreliable that achieving a fully functional unit appeared impossible. While earlier CDC models such as the 6600 and 7600 had encountered similar difficulties, Cray determined that the 8600's design was fundamentally flawed and recommended a complete overhaul to CEO William Norris. Norris rejected this proposal due to financial concerns, insisting that Cray proceed with the existing problematic design.
Disheartened, Cray departed from CDC in 1972 to establish Cray Research, concentrating on single-processor systems because he doubted that current software could effectively utilize multiprocessing—a perspective shaped by the disappointing performance of the ILLIAC IV. The split was on good terms; Norris and CDC personnel invested in Cray Computer's initial public offering, which yielded positive returns.
Following Cray's departure, CDC continued development of the 8600 under Jim Thornton's leadership. By 1974, reliability issues persisted. Simultaneously, Thornton's rival project, the STAR, entered production but demonstrated poor performance in practical applications. The launch of the Cray-1 in 1976 quickly displaced CDC from the supercomputer industry. CDC's subsequent effort to regain market presence with the ETA-10 during the 1980s also ended in commercial failure.