CDC STAR-100
| Company type | Product |
|---|---|
| Industry | Supercomputing, Computer hardware |
| Predecessor | CDC 7600 |
| Founded | 1974 |
| Founder | Control Data Corporation |
| Headquarters | Arden Hills, Minnesota , United States |
| Products | Vector supercomputer |
| Parent | Control Data Corporation |
The CDC STAR-100 represents a vector supercomputer engineered and sold by Control Data Corporation (CDC). It was among the pioneering systems to implement a vector processor aimed at boosting computational efficiency for scientific workloads, and it was the first supercomputer to utilize integrated circuit technology alongside a memory capacity of one million words.
The designation STAR merges the concepts of STrings (binary digit sequences) and ARrays. The suffix "100" signifies its maximum processing capability of 100 million floating-point operations per second (MFLOPS), exceeding the earlier CDC 7600, which reached a peak of 36 MFLOPS but generally operated near 10 MFLOPS.
Conceived as a response to a mid-1960s request from Lawrence Livermore National Laboratory (LLNL) for a high-performance computer within their leasing budget, the STAR-100 was introduced publicly in the early 1970s. On August 17, 1971, CDC revealed that General Motors had placed the initial commercial order.
Although its design was innovative, certain core architectural decisions resulted in performance levels below expectations once it entered commercial use in 1974. This performance gap played a role in CDC losing its leading position following the debut of the Cray-1 in 1975. Ultimately, only three STAR-100 machines were delivered—two units to LLNL and one to NASA Langley Research Center.
Description
The STAR employed a 64-bit architecture encompassing 195 instructions, including 65 specialized vector instructions inspired by the APL programming language, particularly its "control vectors" (vector masks) and vector permutation capabilities.
Vector instructions processed data stored sequentially in virtual memory and supplied an arithmetic pipeline, allowing a single command to add vectors with lengths up to 65,535 elements. Vector operands were retrieved in 512-bit blocks (superwords), which helped minimize memory latency. The CPU could preemptively fetch subsequent operands during instruction execution, enhancing overall throughput despite unchanged individual instruction completion durations.
Numerous instructions were intricate, especially vector macro instructions that executed operations typically requiring multiple simpler commands. This complexity was handled through microcode routines.
The main memory consisted of 65,536 superwords of 512 bits each and featured 32-way interleaving to support pipelined access. Constructed from core memory with an access time of 1.28 μs, it connected via a 512-bit bus managed by the storage access controller (SAC). The stream unit served as the control center, overseeing instruction fetching and decoding, initiating memory accesses, and controlling execution flow, aided by two read buffers and one write buffer for streaming data.
Two 64-bit arithmetic pipelines were integrated: one containing floating point adder and multiplier units, and another dedicated to all scalar instructions plus floating point division. These pipelines were governed by microcode. The system could partition floating point pipelines into four separate 32-bit units, effectively doubling peak throughput to 100 MFLOPS at lower precision.
Input/output processors relieved the CPU of I/O responsibilities. Each I/O processor was a 16-bit minicomputer equipped with its own core memory of 65,536 words and shared a 128-bit data bus connected to the SAC.
Real-world performance, users and impact
In practical applications, the STAR-100's actual performance fell considerably short of its theoretical capabilities. Its "memory-to-memory" vector instructions experienced lengthy initialization delays due to deep pipeline stages and a slower clock cycle (40 ns compared to 27.5 ns on the 7600). As a result, the STAR only surpassed the 7600's speed when processing vector lengths longer than approximately 50 elements; shorter loops suffered from setup overheads that outweighed any performance improvements.
When introduced in 1974, the STAR's overall performance failed to meet user expectations. The majority of software could not be fully vectorized because data dependencies necessitated clearing the pipeline before reuse, leading to substantial setup penalties. Furthermore, scalar processing speed was reduced to prioritize vector operations, which caused a decline in general scalar execution efficiency (refer to Amdahl's Law).
Two STAR-100 machines were installed at Lawrence Livermore National Laboratory, and one was delivered to NASA's Langley Research Center. At LLNL, the STACKLIB library was created to simulate STAR vector functions on the 7600; executing programs with STACKLIB on the 7600 resulted in better performance than previously achieved, underscoring the STAR's relative inefficiency.
Commercially, the STAR-100 did not meet success expectations. Its lead architect, James E. Thornton, departed CDC to establish Network Systems Corporation. An enhanced design emerged in 1979 as the Cyber 203, followed by the Cyber 205 in 1980; however, by this time, Cray Research had introduced significantly more advanced systems. The STAR's shortcomings contributed to CDC losing its leading position in the supercomputer sector and led to the creation of ETA Systems in 1983.
Installations
A total of five CDC STAR-100 units were manufactured and distributed beginning in 1974:
- Control Data Corporation, Arden Hills, MN (2)
- Lawrence Livermore National Laboratory (2)
- NASA Langley
Further reading
- R.G. Hintz and D.P. Tate, "Control Data STAR-100 processor design," Proc. Compcon, 1972, pp. 1–4.
External links
- Neil R. Lincoln with 18 Control Data Corporation (CDC) engineers on computer architecture and design, Charles Babbage Institute, University of Minnesota. Participants include Robert Moe, Wayne Specker, Dennis Grinna, Tom Rowan, Maurice Hutson, Curt Alexander, Don Pagelkopf, Maris Bergmanis, Dolan Toth, Chuck Hawley, Larry Krueger, Mike Pavlov, Dave Resnick, Howard Krohn, Bill Bhend, Kent Steiner, Raymon Kort, and Neil R. Lincoln. Topics cover CDC 1604, CDC 6600, CDC 7600, CDC 8600, CDC STAR-100 and Seymour Cray.