Skip to content

Gene Amdahl and the System/360

Abstract

Gene Amdahl was the chief architect of the IBM System/360, the computer family IBM bet the company on in 1964. With Gerrit Blaauw and Fred Brooks he designed the architecture that unified IBM’s incompatible product lines into one compatible family, proved that binary compatibility across machines of different speeds was achievable, and created the platform on which enterprise computing has run for six decades. He then left IBM, frustrated by bureaucracy, proved that the architecture he had created could be built cheaper than IBM built it, and gave antitrust lawyers more ammunition than they knew what to do with. Along the way, in a short 1967 conference paper, he made the argument now called Amdahl’s Law (the limit that the serial part of a program puts on parallel speedup).

IBM System 360
IBM System/360 Model 30. Image: ArnoldReinhold, CC BY-SA 3.0, via Wikimedia Commons.

From South Dakota to IBM

Gene Myron Amdahl was born on November 16, 1922, in Flandreau, South Dakota, to immigrant parents of Norwegian and Swedish descent who farmed the plains. He studied at South Dakota State College, interrupted by service in the United States Navy during World War II, took a degree in engineering physics in 1948, and went on to graduate work at the University of Wisconsin-Madison. He completed a PhD in theoretical physics in 1952. For his dissertation he designed a computer, the WISC (Wisconsin Integrally Synchronized Computer), about 700 valves and a magnetic drum, which was completed in 1954; the work convinced him that computing was the right field.

He joined IBM in June 1952 as a young engineer at the company’s Poughkeepsie, New York laboratory. Amdahl was assigned to work on one of the most important machines of the early decade: the IBM 704.

The 704, introduced in 1954, was the first mass-produced computer to use floating-point hardware, instructions that operated directly on real numbers in scientific notation, rather than requiring programmers to manage decimal scaling manually. The 704 also had index registers, which allowed systematic traversal of arrays without rewriting addresses by hand. It was the machine for which John Backus and his team at IBM developed FORTRAN, the first widely used high-level programming language, which compiled to efficient 704 machine code and made scientific programming accessible to researchers who were not programmers.

Amdahl followed the 704 with the 709 and early design work on the project that became the IBM 7030 Stretch, the transistorized supercomputer delivered to Los Alamos National Laboratory in 1961. Stretch introduced lookahead execution, the 8-bit byte and a 64-bit word, ideas that fed into IBM’s later architecture work; it also missed its performance targets and was sold at a reduced price. Amdahl was gone before it was finished: frustrated with IBM’s bureaucracy, he left in December 1955 for two young California firms, Ramo-Wooldridge and Aeronutronic, and returned to IBM in September 1960.

The System/360: One Architecture for Everything

By 1961, IBM’s product line had become a serious strategic problem. Over the previous decade, the company had produced a dozen incompatible computer families: the 700 series for scientific computation, the 1400 series for business data processing, the 1620 for scientific work at smaller institutions, the 7090/7094 for advanced scientific computing. Each family had its own instruction set, its own operating system software, its own peripheral devices. A program written for an IBM 1401 could not run on an IBM 7090. A customer who grew out of one machine had to rewrite all their software to move to the next.

This incompatibility was not accidental (it was the product of separate engineering teams working independently) but it was strategically incoherent. IBM was in effect competing against itself. Each upgrade sale required convincing a customer to rewrite their entire software inventory. And competitors were beginning to offer software compatibility as a feature.

The plan came out of IBM’s SPREAD task force of 1961; the architecture was worked out in 1962 and 1963 by Amdahl as chief architect, Gerrit Blaauw and Fred Brooks. It was a single architecture (a precisely defined set of instructions, addressing modes, data types, and register conventions) that would be implemented across a range of machines from the cheapest to the most powerful. Any program written for the cheapest System/360 model would run without modification on the most expensive one. The architecture would be stable; customers would not need to rewrite software when they upgraded hardware.

The technical challenge was implementing this compatibility across a wide range of hardware performance. A fast machine and a slow machine had to run the same instruction set, but the fast machine used completely different hardware. The main tool was microprogramming: implementing the instruction set in firmware, as a sequence of simpler micro-operations, so that very different hardware could implement the same architectural behavior. Slow, narrow machines such as the Model 30 took many micro-operations per instruction; the fastest models were wired directly in hardware. All produced identical results.

Info

Microprogramming, the technique of implementing a complex instruction set by breaking each instruction into simpler micro-operations executed by a simpler underlying hardware machine, was invented by Maurice Wilkes at Cambridge in 1951. The System/360’s achievement was applying it at commercial scale to create the first fully compatible family of computers. The concept of an instruction set architecture (ISA) as a stable interface between hardware and software, allowing hardware to change without forcing software rewrites, is one of the most important abstractions in computing history. It descends directly from the System/360 project.

The IBM System/360 was announced on April 7, 1964; according to IBM’s press release, more than 100,000 businessmen attended announcement meetings in 165 American cities that day. IBM spent about $5 billion on the project over four years, more than its annual revenue. The system offered six initial processor models, with purchase prices from $133,000 to $5.5 million (or $2,700 to $115,000 a month to rent). All models ran the same software.

Fred Brooks, who managed the operating system software development for the System/360, later wrote The Mythical Man-Month (1975) partly from the lessons of that software effort, the OS/360 project had consumed more resources and taken longer than any estimate, and had taught Brooks deeply uncomfortable truths about software engineering and the difficulty of managing large programming teams.

The System/360’s Commercial Success

The System/360 succeeded beyond IBM’s forecasts: more than a thousand were ordered in the first month. Customers who had been considering competitive machines from Burroughs, Honeywell, or Control Data moved to the System/360, buying not just hardware but the implicit promise of software continuity, the ability to upgrade hardware for decades without rewriting software.

The architecture survived not years but decades. The System/360 was extended into the System/370 (1970), which gained virtual memory in 1972. The 370 became the 303x and 308x families, then the 3090, then the System/390, then the IBM Z series, the modern mainframe line. The fundamental register set, instruction set, and data type definitions of the 1964 System/360 are still present in the IBM Z architecture of 2024, running financial transactions for most of the world’s largest banks. The architecture Amdahl designed has been in continuous production and use for sixty years, which may be longer than any other computer architecture in history.

Amdahl’s Law

In 1967, at the AFIPS Spring Joint Computer Conference in Atlantic City, Amdahl presented a three-page paper whose argument became known as Amdahl’s Law. The paper contains no formula; the equation below was written down later by others.

The observation was a response to advocates of parallel array computers such as Daniel Slotnick, who argued that connecting many processors would multiply computing power proportionally. Amdahl’s counterargument was that every real program contains some fraction of computation that is inherently sequential: that must be completed before the next step can begin. If even a small fraction of a program is sequential, the sequential fraction imposes a ceiling on the speedup that parallelism can achieve.

If a fraction p of a program can be parallelized and the remaining fraction (1−p) must run sequentially, then the maximum speedup from N processors is:

S(N) = 1 / ((1-p) + p/N)

As N grows toward infinity, the speedup approaches 1/(1-p). A program that is 95% parallelizable achieves at most 20-fold speedup, regardless of how many processors are applied to it. A program that is 99% parallelizable achieves at most 100-fold speedup. The 1% or 5% that cannot be parallelized is the bottleneck, and no amount of additional hardware changes it.

The law was stated as a critique of unrealistic expectations, not as a fundamental physical limit. John Gustafson, in 1988, argued that Amdahl had framed the problem incorrectly: as more processors become available, programmers typically scale up the problem they tackle, not keep the problem size fixed. Under Gustafson’s framing, the speedup scales more favorably. The tension between Amdahl’s and Gustafson’s perspectives has organized much of parallel computing research for fifty years.

Amdahl Corporation: Proving IBM Overcharged

Made an IBM Fellow in 1965, Amdahl ran IBM’s Advanced Computing Systems laboratory in Menlo Park in the late 1960s on a high-performance computer project, which was cancelled in May 1969. In September 1970 he left IBM again, after his proposals for a new high-end machine were rejected. He had spent about fourteen years at IBM in two spells, had been responsible for one of the company’s most important products, and left with the conviction that IBM was substantially overcharging customers for its hardware.

On 19 October 1970 he founded Amdahl Corporation in Sunnyvale, California, at first with the finance specialist Ray Williams and two secretaries, with the intention of building IBM-compatible mainframes, machines that ran IBM’s System/360 instruction set and all System/360 software, but were built with newer technology and more efficient circuit design, delivering better performance at lower cost.

Funding was difficult. IBM’s reputation was so powerful that investors were skeptical that any competitor could credibly challenge it. Amdahl spent most of the first five years seeking capital rather than building computers. In a recession year he raised $2 million of the $44 million he needed. The breakthrough came through Fujitsu, the Japanese technology company, which put in $5 million in 1971, followed in 1972 by Heinz Nixdorf with $6 million; their entry brought in further investors. The arrangement gave Fujitsu access to Amdahl’s design work, and later made Fujitsu the dominant shareholder; it gave Amdahl the capital to finish his machine.

The Amdahl 470V/6 shipped in 1975. Compared with the IBM System/370 Model 168 it replaced, it was about twice as fast and about 10 percent cheaper, and it was air-cooled, using advanced integrated circuits. Fifty systems had been sold by spring 1977, when the company, listed since 1976, earned $48 million on $321 million revenue; the 470V/7 of 1978, its answer to the IBM 3033, sold about a hundred in a year. Nixdorf sold his stake at a profit in 1977, leaving Fujitsu as the dominant shareholder. IBM’s response was immediate: price cuts, new lease terms designed to penalize customers who switched mid-lease, and quiet pressure on customers through IBM salespeople. The U.S. Department of Justice had already filed an antitrust suit against IBM in 1969.

Info

The IBM antitrust case, United States v. IBM, was filed in 1969 and dragged on until 1982, when the Reagan administration dropped it as “without merit.” The case occupied enormous resources on both sides (IBM deployed hundreds of lawyers and produced millions of pages of documents) and its resolution (or abandonment) left the question of IBM’s competitive practices legally unanswered. Amdahl Corporation’s success in the plug-compatible mainframe market was taken, by those outside IBM, as practical evidence that IBM had indeed been overcharging for computing power.

Amdahl Corporation was profitable through the 1980s, reaching approximately $2 billion in annual revenues at its peak. The strategy depended on IBM maintaining a stable architecture that Amdahl could be compatible with, and on IBM’s pricing remaining high enough to leave a margin for a competitor. Both conditions held through the 1980s.

Trilogy Systems and the Failed Bet on Wafer-Scale Integration

In 1979, Amdahl left his own company, chiefly because of Fujitsu’s dominant influence, and set out on a more ambitious technical plan: Trilogy Systems, which he founded together with his son Carl Amdahl and Clifford Madden. Trilogy’s goal was to build extremely high-performance computers using wafer-scale integration, rather than cutting a silicon wafer into individual chips, build an entire circuit on a single wafer, with multiple redundant copies of each functional unit and on-wafer routing to work around manufacturing defects.

Warning

Wafer-scale integration had been an engineering dream since the early 1970s. The economics were appealing: a single wafer-scale circuit would eliminate the packaging and interconnect overhead of chip-to-chip connections, potentially enabling much faster systems. The practical problem was yield. Even the best semiconductor processes have some fraction of defective transistors per unit area. On a small chip, you either get a working chip or you discard it; the good chips far outnumber the defective ones. On a wafer-scale device, no circuit the size of a wafer is free of defects, and the complexity of routing around defects reliably enough to produce a functioning system was never solved. Trilogy raised about $230 million, a record for a start-up at the time, plus $60 million in a 1983 public offering; a winter storm damaged its fabrication plant while it was being built, and defects did the rest. In mid-1984 it gave up on manufacturing wafer-scale computers and Amdahl stepped down as chief executive; in 1985 Trilogy used its remaining money to merge with the minicomputer maker Elxsi, which Amdahl left in 1989.

Legacy

His next company, Andor International (1987), went bankrupt in 1995, and his last, Commercial Data Servers, founded in 1996, stopped making hardware in 1999. Amdahl Corporation, with about $2 billion in revenue in each of 1988 and 1989, was taken over completely by Fujitsu in 1997. Amdahl died of pneumonia on November 10, 2015, at ninety-two, in Palo Alto. He had spent sixty years building some of the most capable and influential computers of each era (the 704, Stretch, the System/360, the 470) and had contributed a conceptual tool that every parallel systems designer still uses.

The System/360 architecture continues in IBM Z mainframes. Its stability across six decades is both its achievement and its limitation: the 1964 design had no virtual memory and no networking, and the successive extensions have not always been elegant. But it has survived, and it works, and the investment customers made in System/360 software has been preserved across generations of hardware.

Amdahl’s Law is taught in every computer architecture course. The debate it triggered, about the limits of parallelism, about how to think about the ratio of serial to parallel work, about whether the right measure is speedup at fixed problem size or throughput at variable problem size, remains live. As computing has moved from single fast processors to massively parallel GPU clusters running machine learning workloads, the question of which fraction of computation is inherently sequential has become practically urgent again, sixty years after Amdahl asked it.

The System/360 project and its role in IBM’s dominance is covered in The Supercomputer Era. The semiconductor context of Amdahl’s career is in The Integrated Circuit Revolution.

📚 Sources