IBM Advanced Computer Systems project

IBM Advanced Computing Systems (ACS)
Industry	Supercomputing, Computer hardware
Founded	1965
Defunct	1969
Fate	Project cancelled; technologies influenced later IBM and other architectures
Headquarters	Sunnyvale, California, United States
Key people	Max Paley, Jack Bertram, John Cocke, Frances Allen, Herb Schorr, Lynn Conway, Gene Amdahl
Products	ACS-1, ACS-360 supercomputers (never completed)
Parent	IBM

The ACS-1 and ACS-360 were supercomputing systems developed by IBM as part of the Advanced Computing Systems initiative from 1965 to 1969. While these machines never reached completion or commercial production, the project introduced organizational strategies and architectural breakthroughs that have become foundational in the realm of high-performance computers. The innovations from this endeavor influenced the design of the IBM RS/6000 series and played a role in shaping the Explicitly Parallel Instruction Computing (EPIC) architecture implemented in Intel and HP’s Itanium processors.

Following the termination of the ACS project, numerous engineers chose not to relocate back east and instead established MASCOR in 1970, a venture that ultimately failed due to insufficient funding. Among them, Gene Amdahl founded Amdahl Corporation, which produced IBM-compatible mainframe systems utilizing ACS-developed ECL technology. The Amdahl 470V/6 notably delivered superior performance at a lower cost compared to IBM’s top-tier models.

IBM and CDC

IBM’s inaugural supercomputer, the IBM 7030 Stretch, debuted in 1961 but was withdrawn after underperforming in practical applications at Los Alamos Scientific Laboratory. Subsequently, IBM initiated two separate projects: Project X, based at IBM Poughkeepsie Laboratory, aimed to achieve speeds 10 to 20 times faster than Stretch, while Project Y, operating out of the IBM Thomas J. Watson Research Center, set an ambitious goal of reaching performance levels 100 times greater.

In 1962, Control Data Corporation (CDC) announced installations at Lawrence Radiation Laboratory and introduced a powerful new system, the CDC 6600, which was publicly revealed in 1963. This development led IBM CEO Thomas J. Watson Jr. to question how CDC, a smaller enterprise, managed to surpass IBM’s capabilities.

By September 1963, IBM strengthened its high-end New Product Line (NPL). Project X adopted the NPL instruction set and evolved into the Model 92 (later renamed Model 91), becoming part of the successful Model 90 product family. Meanwhile, Project Y remained independent from NPL, focusing on scientific computing under Jack Bertram’s Experimental Computers and Programming Group. Significant contributors included John Cocke, Frances Allen, Brian Randell, Herb Schorr—who crafted the initial instruction set—and Lynn Conway, who was brought on board for system simulation tasks.

Move to California

Although the System/360 achieved immediate market success, production challenges delayed shipments and demanded considerable company resources. Concurrently, CDC announced a new machine boasting tenfold performance improvements over the 6600. Watson doubted whether the System/360 instruction set was suitable for such an advanced design and feared further delays given ongoing System/360 difficulties. In early 1965, he authorized establishing a new division in California to be closer to weapons laboratory clients. IBM acquired a facility in Sunnyvale, California to house the IBM Advanced Computing Systems (ACS) laboratory under Max Paley’s leadership.

During a steering committee meeting in August 1965, Paley, Bertram, and Schorr unveiled their design: adopting a standard word length of 48 bits tailored for scientific workloads; targeting a clock cycle time of 10 nanoseconds—approximately ten times faster than the CDC 6600—with six or seven internal cycles per clock tick. The arithmetic logic units (ALUs) featured pipelining capable of issuing multiple instructions simultaneously. Branch execution was optimized through a buffer mechanism that processed both potential branch paths.

Harwood Kolsky evaluated competing architectures while Gene Amdahl and Chen Tze-chiang discussed their efforts on the high-end System/360 Model 92. Kolsky brought experience from Los Alamos before joining Stretch; Amdahl had initially left IBM after being overlooked for Stretch development but returned in 1960 to contribute to Project X. In late 1964, Amdahl accepted a teaching role at Stanford but was appointed an IBM Fellow in January 1965 for his work on Model 92. As an IBM Fellow, he accepted Bob Evans’ invitation to join ACS.

At this preliminary phase, Amdahl advocated for ACS compatibility with System/360 architecture as a more viable approach. Although this might slightly reduce speed due to having fewer registers (sixteen 32-bit versus thirty-two 48-bit), compatibility would offer Model 92 users an upgrade path toward enhanced performance while enabling them to utilize existing System/360 software and compiler infrastructure.

Design development progresses

By early 1966, the Project Y blueprint was finalized and designated as ACS-1, with the primary modification being the elimination of the 192-bit extended floating-point format. That same year, a new 38,000-square-foot building was constructed at 2800 Sand Hill Road in Menlo Park, California, close to the Stanford Linear Accelerator, and the project team moved there toward the end of the year. A significant change in design philosophy took place: initially, the compiler transferred instructions from large core memory or thin film memory into a smaller static RAM cache within the CPU. However, Schorr and Dick Arnold determined this approach was unworkable and redesigned it into a single-level hardware cache of either 32 or 64 kWords.

Another breakthrough was the introduction of dynamic instruction scheduling (DIS). The arithmetic logic unit (ALU) and indexing units featured six-slot buffers that could select two instructions for out-of-order execution, allowing queued instructions to proceed while earlier ones awaited required data. The results were then reordered in memory to preserve correct machine code sequencing. Lynn Conway, hired to develop a software simulation, devised a bit-matrix system to monitor instruction readiness.

Utilizing this simulator, Conway benchmarked demanding workloads against the IBM 7090, CDC 6600, and S/360 Model 91. Relative to the 7090, ACS-1 executed the Lagrangian Hydrodynamics Calculation (LHC) approximately 2,500 times faster and surpassed it by nearly 1,300 times on Neutron Diffusion (ND) code; it was roughly 60 times quicker than the CDC 6600.

Advanced compiler development for ACS-1 was spearheaded by Allen, Cocke, and Jim Beatty, emphasizing aggressive optimization techniques. These included loop unwinding, instruction scheduling around basic blocks, and differentiating between code-specific and platform-specific optimizations. The compiler supported both a PL/1 front-end as well as an enhanced Fortran IV.

In a November 1967 progress review, Herb Schorr outlined a delivery schedule aiming for initial shipment in 1971. The plan anticipated over 100,000 lines of Fortran and assembly language for the operating system, with nearly 70,000 lines devoted to compilers, assembler, and libraries. Software development expenses were projected at $15 million.

Design "shootout"

Amdahl advocated for creating a version compatible with System/360 architecture. In January 1967, Ralph L. Palmer assigned John Backus, Robert Creasy, and Harwood Kolsky to assess both the project and Amdahl’s proposal. Kolsky judged System/360 compatibility as too challenging and noted that since ACS targeted the CDC 6600 market segment, compatibility with the CDC 6600 would be more pertinent. The following month, Amdahl renewed his push for System/360 compatibility citing marketing advantages.

This viewpoint increasingly clashed with Bertram’s approach, who effectively "quarantined" Amdahl by restricting his interactions. Simultaneously, circuit designer John Earle was removed from the primary team due to interpersonal conflicts and reassigned to work with Amdahl—apparently as a form of reprimand. This move backfired when Amdahl persuaded Earle that a System/360-compatible design was achievable, resulting in the development of the Amdahl-Earle Computer (AEC/360). Drawing on ACS-1 concepts, they engineered a system that was somewhat slower but cost approximately 75% less to manufacture and utilized only about 90,000 gates compared to ACS-1’s 270,000. The reduction in gate count was mainly due to fewer registers but compensated by a faster clock speed of 8 nanoseconds enabled through a simplified design.

In December 1967, Kolsky met with Amdahl to review design specifics. Amdahl also began internally sharing his design at IBM’s System Development Division in New York, prompting vice president Erich Bloch to initiate a formal review. In response, the ACS team proposed a redesign cutting gate count from 270,000 down to 200,000 with minimal performance degradation—indicating possible overengineering.

Carl Conti from IBM Poughkeepsie led the March 1968 evaluation. Amdahl presented hand-calculated performance projections: on integer benchmarks, AEC/360 could be up to five times faster than ACS-1; however, on floating-point operations it might be up to 2.5 times slower; its simpler branching mechanism could be adapted from ACS designs. Conti noted that ACS’s heavy reliance on compilers for performance might pose risks if ported elsewhere. He estimated AEC would require roughly 108,000 gates—about half the complexity of ACS.

A brief final assessment took place in April. By May, IBM decided to cancel ACS-1 and move forward with AEC/360 (later renamed ACS-360). While Amdahl’s competing design influenced this decision significantly, other factors included his argument that allocating $15 million toward improving System/360 operating systems would benefit IBM’s entire product line more effectively—and NASA’s delivery of a powerful System/360 Model 95 in January 1968. Although ACS would have outpaced Model 95 substantially, IBM contemplated exiting the supercomputer market.

Looking back, many agree that ACS-1 would have ranked among leading supercomputers of its time. Conway remarked it "would have been the premier supercomputer of the era." The cancellation largely resulted from reliance on unverified cycle count estimates—the simulator had not been updated for the revised instruction set. While Conti accepted Amdahl’s assertion of an 8-nanosecond cycle time, some experts remain skeptical about its feasibility.

Cancellation

Following the departure of most ACS senior leaders, control was assumed by Amdahl. The development of the AEC/360 proceeded largely as intended, with significant modifications such as the implementation of generalized register renaming within the out-of-order execution framework and enhancements to the branch prediction mechanism for the 360 instruction set.

Amdahl concluded that the machine's sales would not be financially viable, presenting a considerable threat to IBM. Launching a premium system destined to incur losses risked being perceived as anti-competitive, potentially undermining markets held by competitors like CDC. IBM had encountered similar issues with Stretch but managed to balance expenses by leveraging research and development efforts across multiple projects. To boost profitability, Amdahl suggested manufacturing three variants: the original ACS/360, a smaller model delivering one-third of its performance, and an even more compact version at one-ninth performance—still outperforming any IBM machine at the time. This plan was ultimately declined.

In May 1969, IBM's leadership chose to terminate the project, reportedly upon Amdahl's recommendation. Initially designed to rival CDC’s high-speed machines, the initiative had extended nearly ten years without producing a marketable product. Amdahl later explained that cancellation was primarily due to disruption in IBM’s pricing strategy—machines exceeding a certain performance threshold were expected to operate at a loss, and competitively pricing the ACS/360 would necessitate price reductions across other product lines. He also hinted at speculation that the project might have been deliberately set up for failure so its innovations could be repurposed and research expenses written off.

Shortly after termination, in August 1969, IBM unveiled the IBM System/360 Model 195, a redesigned version of the Model 91 utilizing integrated circuit technology that doubled the speed of the then-leading Model 85. To further address high-performance computing needs, IBM launched a vector processing task force based in Poughkeepsie.

After ACS was discontinued, many engineers chose to remain in California rather than return to IBM’s research center in New York. Some joined IBM’s hard drive manufacturing facility in San Jose, California, while others established Multi Access System Corp (MASCOR), which ceased operations within months. Amdahl resigned in September 1970 and founded Amdahl Corporation to produce IBM 360-compatible systems, introducing the Amdahl 470/6 in 1975. This company became a significant supplier of IBM-compatible machines through the 1970s and 1980s, securing over 20% of the market during that timeframe.

Influence

Although neither ACS-1 nor ACS-360 reached production, IBM’s Advanced Computing Systems team pioneered numerous architectural breakthroughs that have influenced contemporary computer design, especially in RISC CPU methodologies:

Significant minimization of logic gate levels within pipeline stages to reduce cycle duration
Close integration between processor and memory systems
Implementation of cache memory paired with efficient I/O pathways
Advanced compiler optimization strategies
Operating systems supporting virtual memory
Multiple instruction decode and issue capabilities (a pioneering achievement)
Introduction of branch target buffer technology (a first)
IBM's initial use of hardware multithreading
Dynamic instruction scheduling enabling out-of-order execution
Hardware-based register renaming
Use of instruction predication
Application of Level-sensitive scan design by IBM
Development of fixed-head hard disks
Air-cooled high-speed large-scale integration (LSI) circuits
Sophisticated simulation tools for design validation