In the mid-1990s, IBM was facing an unfamiliar problem. The company had built its reputation on producing a long line of incredibly powerful computers, from machines like the 7030 Stretch, which reigned as the world’s fastest in the early 1960s, to RISC-based machines and the IBM PC. But by 1996, IBM ranked fifth out of five companies in the burgeoning Unix server market — primarily due to the speed of its microprocessors. The exploding popularity of the internet significantly increased the demand for faster servers, and IBM was caught flat-footed. Incremental improvements wouldn’t solve the issue. An entirely new approach to processor architecture was in order.
So, a handpicked team of 250 systems and chip designers, software architects, researchers and semiconductor engineers set out to reinvent the microprocessor and much of its supporting architecture. In the arcane world of CPU design, IBM engineers and scientists had to break new ground in pipelining — the process of accumulating instruction from the processor in an orderly fashion, which keeps all portions of the processor occupied and increases the amount of useful work the processor can do at a given time. Engineering advances were also necessary in multithreading, a CPU feature that enables two or more instruction threads to execute independently while sharing the same resources. This allows multiple concurrent tasks to be performed within a single process. The hope was to leapfrog competitors with a wholly new server system that would offer twice the performance at half the cost.
The project took nearly five years to complete, but the result was a game changer. The team successfully grafted two high-performance microprocessor cores on a single silicon chip — an industry first — and developed a novel computer architecture to support it. The dual-core processor enabled multiple programs to run simultaneously, or a single program to run faster and more efficiently. Dubbed IBM Power4 (POWER was originally an acronym for “Performance Optimization With Enhanced RISC”), the new processor served as the brains for an IBM eServer pSeries server called IBM Regatta. It met the goal of more than doubling the speed of the nearest competitor at half the cost. Microprocessor research company Cahners In-Stat/MDR awarded the Power4-based system its Analysts’ Choice Award for Best Workstation/Server Processor of 2001.
In 1996, Vijay Lund was working on mainframes when he was tasked with finding a way for IBM to regain its lead in Unix servers. Lund deduced that the only route to dominance would be overhauling the underlying computer architecture with a far more powerful microprocessor and vastly improved design. In a meeting with IBM Senior Vice President Nick Donofrio, Lund recommended that IBM build its own microprocessor to lock in the new system’s advantages.
“I was thinking about the long term,” recalled Lund. “If we did it ourselves, I knew we would get it right, and our solution would be a starting point for a series of competitive systems — not just a one-time step forward.”
Lund assembled an A-team of developers from IBM locations in the US, Canada and Germany. The company also established an alliance with Hitachi, which would build part of the processor’s on-chip memory, called a Level 3 cache. In 1996, complex developments of this magnitude usually fell apart unless the key people worked under the same roof. To beat those odds, the team perfected early versions of distance collaboration tools, such as remote desktop sharing. Such technology is commonplace today, but it was a novel approach for the 1990s. Lund also paired risk-takers with seasoned realists to guard against erring on the side of an especially radical or conservative approach.
The creative friction paid off. After nearly five years of intense effort, the team decided to combine two processors on a single chip — something that had never been done before. Power4’s industry-first “server on a chip” was manufactured with a proprietary process that enabled the chip to process data much faster.
Traditionally, designers improved chip speed by packing ever more circuits together on a piece of silicon. However, when in close-enough proximity, circuits generate interference that hampers operations. IBM developed a recipe for building chips with a low-interference material that forms a better seal around the chip’s wiring. This greatly reduced the interference that would normally have occurred with two processors so close together on a single chip.
This breakthrough design in architecture and materials engineering enabled the two processors to work together at a very high bandwidth with large on-chip memories and high-speed input/output channels. Four of these new microprocessors working together as a powerful 8-way module established an industry standard and produced a then-record clock speed of 1.3 gigahertz.
As IBM’s J. M. Tendler and four co-authors wrote in the IBM Journal of Research & Development, “In the ongoing debate between the ‘speed demons’ (high clock rate) and the ‘brainiacs’ (more complex design but a higher instructions-per-cycle rate), IBM Unix-based systems have traditionally been in the brainiac camp. With the Power4, IBM opted to also embrace the speed-demon approach.”
In October 2001, IBM launched what was then the world’s fastest server, code-named Regatta, around Power4. The breakthrough processor design enabled Regatta to conserve energy and dramatically outperform servers with twice as many processors. Power4 was an instant hit with customers and reviewers, and vaulted IBM back into the lead of the Unix server segment. “Everything about IBM’s Power4 processor is amazing,” gushed a review in PC Magazine. “Amazing technology, amazing size (680 million transistors), amazing power requirements, amazing performance.”
The new system-level architecture spawned an entire Power series, which became the heart of IBM’s line of industry-leading servers for decades to come. In 2005, more than a dozen companies, including Sony and IBM, formed Power.org, an open-standards community focused on accelerating Power chip collaboration. Power microprocessors became the core of everything from game consoles to powerful supercomputers. Watson, the computer that defeated human opponents on the TV quiz show Jeopardy!, was built on Power7 technology.
In 2021, IBM introduced the Power10 processor, designed specifically for hybrid cloud environments. It’s the heart of the IBM Power E1080 server, which can deliver up to 30% more performance per core, while consuming 33% less energy, compared with the previous generation of IBM servers — enabling customers to consolidate workloads and save on hardware and software costs.
The continuously high level of microprocessor performance is further proof that IBM’s breakthrough design in the late 1990s has had staying ... power.
With ‘Stretch,’ IBM entered the age of supercomputing
A USD 1,500 open-architecture machine became an industry standard and brought computing to the masses
An inspired researcher overcame a nagging technical problem to pioneer massive opportunity in semiconductors