Microprocessors don’t get much respect any more. Mind you, they’re at the heart of every desktop, every server, every router and many other parts of the company network. But somewhere along the line, the processor lost its power to dictate customer decisions.
It wasn’t always that way. Back in the halcyon days of the processor, every system architecture was new and each new processor hoped to dominate the market.
Systems of the 1970s had compute power concentrated in one place, and each increase in CPU speed — from 8MHz to 12MHz, for example — was a major new advance as the fledgling processor industry put Moore’s Law through its first tests.
Back then, vendor lock-in was still a viable business strategy. IBM, Digital, Honeywell, Unisys and their competitors grew by finding interesting ways to link processor, memory, storage devices, services and other elements into profitable long-term customer relationships. This approach engendered significant competition in the microprocessor space, advancing the state of the art with dizzying speed.
As Unix became increasingly important in the late 1980s, mid-range systems presented a paradigm shift as Unix distributions proliferated and each, tied to a specific processor architecture, found its niche. At a time when Intel was still making low-powered desktop CPUs and a server was simply a fast 486, businesses were focused on proprietary RISC-based processors from Silicon Graphics, Sun Microsystems, HP, IBM, Digital, and Unisys.
Somewhere along the line, however, customers began caring less about what was under the system’s hood than what it let them do. Ever-faster 32-bit Intel processors became an increasingly attractive proposition, and with successive generations it became clear that customers were loath to shell out for proprietary RISC systems when cheaper, industry-standard Intel servers would do the trick.
“If you’re different and 5 per cent faster, there’s no value proposition there,” says industry consultant Terry Shannon, who has followed Digital’s strategy since 1975 and observed its senior management trying to fight the Intel juggernaut during the 1990s before eventually signing the Alpha chip’s death warrant. Noting that operating systems such as Tru64 and VMS have been smoothly ported between processor architectures in the past, Shannon says vendors realised early on that vendor lock-in was not a long-term strategy.
“Compaq prudently realised different isn’t necessarily better, and more and more vendors are going to realise it makes no sense to reinvent the wheel.”
The implications of this shift have been quite dramatic, with Intel servers and equivalents now being promoted by RISC vendors whose entire business used to be based around saying why their
systems were better than those based on Intel chips. The only significant distinguishing factors between RISC and Intel chips have rapidly disappeared as Intel’s increasing speed, the introduction of new technology such as hyperthreading, and increasing cache sizes eke out far more bang for the buck than RISC can ever hope to provide.
The second generation of Intel and HP’s 64-bit Itanium chip, whose early shortcomings provided breathing room for RISC competitors, has become the first nail in the coffin for RISC suppliers finding 64-bit capabilities are no longer a differentiator. SGI was one of the first vendors to capitulate to the Intel steamroller, with IBM, HP, Unisys, Digital (now Compaq) and, most recently, Sun, all adding Intel-based systems to their product lines.
It’s not that Intel chips do things that RISC can’t — but it wouldn’t matter even if they did: simple economics dictate that low-volume RISC chips are no longer financially viable in an industry where Intel’s high-volume strategy has pushed prices to levels so low that spending hundreds of millions of dollars to develop a new chip is no longer feasible.
Shannon, for one, estimated Compaq was spending $US150 million a year to develop Alpha, but would sell just 500,000 of the processors in a good year.
“They’re all good chips, but they’re all expensive and there’s a lot of infrastructure that goes around them,” says David Bannon, systems manager with Victorian Partnership for Advanced Computing (VPAC), a Melbourne-based university consortium that began with a 128-processor AlphaServer SC but this year purchased a 194-CPU Linux cluster running standard Intel Xeon processors.
That system’s cost competitiveness drove VPAC to evaluate Itanium 2 chips when it recently specced a new Linux-based supercomputer for a collaborative design project with Holden. The need for massive amounts of memory and a large application space demanded the sophistication of a 64-bit architecture; while Linux runs on most RISC platforms, assessing Itanium 2 against other 64-bit options quickly produced a clear winner.
“The difference now is that there’s an alternative” to expensive RISC chips, Bannon said. “If we hadn’t the capability of buying 64-bit chips at almost commodity prices, we would have had to put in a [Compaq] Alpha or [IBM] Power4 based system. If you’ve made the decision to go with a commodity operating system, perhaps you should stick with a commodity processor.”
Although support for Intel-based systems continues to grow, rival companies are still pursuing roadmaps designed to extend RISC architectures as long as possible. Part of this was born out of necessity — HP, for example, ended up reviving PA-8000 design when it became clear the original Itanium wasn’t going to meet expectations.
For others, the urge to develop RISC comes from anti-Wintel sentiment (in the case of Sun) or a desire to build new markets by expanding into new sectors, such as IBM is hoping to do by building its Power5 chip into Sony’s PlayStation 3. Then there’s AMD, which is hoping its 64-bit Opteron chip can give Itanium 2 a run for its money in the same way that AMD’s technically superior Athlon forced Intel to accelerate its Pentium III and 4 speed roadmap.
Barring some revolutionary change backed by mass-market production capabilities, it’s clear that commodity (read: Intel) processors will continue increasing their share of the server market — and, by extension, the processor market. What was once a mechanism for vendor lock-in has become little more than an engine upon which vendors build the real value-add they need to compete.