Vol 19, Issue 13

March 28, 2005

AMD vs. Intel, Round IX

By Kevin Krewell

Get ready for the next great AMD vs. Intel battle: the dual-core duel of 2005. Is this beginning to sound like the Rocky movie series? And is it getting as worn out as Rocky V? These two companies have been fighting since Intel first tried to nullify the 1981 cross-license agreement that allowed AMD to second-source x86 processors.

There have been cases of cooperation, or détente, but they have been rare. This latest challenge can be traced to Intel's change in its mainstream processor roadmaps, as the company has turned its back on the clock-frequency race and turned to multicore processors as its future direction. The gauntlet was thrown down to AMD: ship dual-core processors or fall behind Intel's marketing and engineering direction. Well, AMD was working on its dual-core processor but had not made a public commitment. Given AMD's limited fab capacity (compared with that of Intel), the company doesn't want to make a broad commitment to dual-core processors, as that would, in essence, halve AMD's capacity to ship processors.

A Quick History Lesson

For those new to this battle, often characterized as a David vs. Goliath matchup, I'll try to summarize the action to date. In the late 1970s and early '80s, Intel had a cross-license agreement and had signed AMD as a second source for various chips, including the x86 processors. In exchange, AMD committed to provide Intel with the rights to second-source some of its support chips. This was very common in the '70s and '80s, as many companies and government contracts required multiple sources. Multiple sources allowed competitive bidding and alternative sources in case one company had a manufacturing problem, such as a yield bust (much more common back then).

The problem for Intel was that AMD was not happy being a docile second source. AMD was much more aggressive and, to Intel's consternation, produced a significantly faster 286 (16MHz vs. 12.5MHz) than Intel was capable of. This situation may well be the origin of the clock-frequency war that AMD and Intel continued to fight for almost two decades; it could be called Round 1 of that fight. Intel then introduced the 80386 and decided to purposefully slow AMD's progress by refusing to hand over the design to AMD. Intel had begun to build its now famous "copy-exact" manufacturing methodology and convinced IBM and other PC manufacturers that traditional second sources weren't required when Intel had multiple fabs in multiple locations, and that Intel was leading in the new-product development essential to creating leading-edge PCs.

Intel even embarked on an ad campaign that denigrated its own 286 processor as old technology and promoted the 32-bit 386 as the future architecture (the ad picture had a big "286" with a big red X through it). AMD, not having the 386 at first, pointed out that there wasn't much 32-bit x86 software available and that Microsoft's Windows 3.1 was a 16-bit operating system. (How ironic that Intel fought so hard against the 64-bit extensions, using almost the same arguments AMD made against the 32-bit extensions!) That approach did not stop the eventual success of the 386 architecture (Round 2).

AMD eventually reversed-engineered the 386 (and later the 486), and once again AMD produced faster (40MHz) versions of the 386 than did Intel (33MHz), although one could argue that Intel had moved on to the 486 when AMD made the faster 386 (Round 2.1). Intel also tried developing a version of the 386, specifically for mobile, with an early form of power management—the 386SL. AMD countered with a less expensive, and faster, version of the 386SX (Round 2.2).

AMD and Intel entered into a series of court battles as Intel withheld the 486 design from AMD as well and sued AMD to prevent shipment of 386s and 486s. AMD was late with the more highly integrated 486 processor (Round 3) as the court battle went back and forth between the two companies. Eventually, AMD won the rights to produce x86 processors and signed a new cross-license agreement with Intel (a brief détente between rounds).

When AMD realized that Intel would not hand over future x86 processor designs, it began its own independent processor design that would eventually become the K5. The K5 was a very ambitious design—too ambitious for the design team, it turned out. The K5 was late and, when it shipped, it was too slow to compete effectively with Intel's Pentium processor (which one AMD engineering manager had earlier derided as only two 486s bolted together). And it was with the K5, and last generation of 486-based processors (which AMD tried to position against Pentium as the Am5K86-PR75) that AMD became involved with the "performance rating" (or PR) system that attempted to show that clock frequency and performance are not synonymous. This occurred because the Pentium processor had a faster clock frequency—even if it had a simpler microarchitecture—and was winning the performance war (Round 4).

With the K5 delayed, AMD had a brand new fab and not enough demand for its product to fill the fab. The solution to AMD's troubles was found in a smaller competitor that was rapidly running out of money but had a new processor design well along: NexGen. The NexGen processor replaced its proprietary bus with a Pentium bus and became the AMD-K6. The 233MHz K6 was the fastest PC processor in 1997 (see MPR 3/31/97-01, "K6 is World's Fastest x86 Chip")—for about three weeks (Round 5). Then Intel launched the Pentium II with clock speeds up to 266MHz (Round 5.1). But at least AMD was back in the game.

Another editorial (see MPR 8/16/04-01, "Who Really Deserves Credit for the New AMD?") describes some of the system battles that took place from the AMD-K6-2 time to today. Intel's Pentium II and Pentium III kept a lead over AMD's K6 family, but AMD had another processor in the works that was even more promising—the K7, later called Athlon. Athlon leapfrogged Intel's Pentium III and was the first PC processor to hit 1GHz. (See MPR 3/13/00-02, "Athlon Wins Race to 1GHz by Hair.") Intel launched the 1GHz Pentium III but had trouble producing it in volume; later that year Intel had to cancel a 1.13GHz speed upgrade (Round 6).

Intel's next processor architecture—the Pentium 4 or NetBurst architecture—put clock frequency at the forefront. It was late, but once it shipped, Intel handily won the clock-speed race (Round 7). However, AMD then changed the ground rules, resurrecting a form of PR to keep the slow Athlon processor competitive, using a mix of recognized benchmarks as the measure of performance (Round 7.1). Once again, AMD had another new architecture in the wings—Hammer. The Hammer architecture became the Athlon 64 and Opteron processors, and AMD took a leadership role, bringing 64-bit extensions, on-chip memory controllers, and glueless multiprocessing to mainstream markets (Round 8). Intel responded by embracing the 64-bit extensions (EM64T) and increasing the processor front-side bus frequency to increase bandwidth and reduce memory latency (Round 8.1). Microsoft's delay of more than a year in shipping the x64 version of Windows XP has certainly helped negate AMD's lead in 64-bit.

That brings us up to Round 9—which I define as the race to dual-core processors. Although both competitors have different designs and different approaches, they share a similar overall strategy: both are implementing first-generation dual-core processors with independent L2 caches and limited cooperation between the cores on power management. AMD has a bit of an edge with the integrated memory-controller crossbar switch, which should allow much faster inter-core coherency traffic than the Pentium 4 front-side bus that the Pentium D will use. Both vendors will ship dual-core processors that run at clock speeds below the fastest comparable single-core processors. Both companies will have to position the dual-core processors as then next step in processor development, even though application software support for multi-core processors will lag behind the hardware.

My concern is whether the companies are moving too quickly to multicore client computers, and that this competitive nature is pushing both companies to move too fast to a technology whose benefits are not clearly defined. Although server application software (such as web services and databases) can use multicore processors, the case for multicore client software is still a work in progress. The initial case for multicore client computing will revolve around multitasking, but Intel and AMD must develop a new set of benchmarks to show the benefits that may not represent the typical PC usage patterns. Can the processor companies convince consumers and business customers that a multicore version of a processor is important technology to adopt? I think the selling of multicore client processors will be an uphill struggle for all but the early adopters. Both companies will still have single-core processors on their roadmap, so if there is resistance to the move (to multicore), the companies have a backup plan.

Spurred on by feedback from editorial board member Jeff Deutsch, I looked back into the MPR archive from 1993 and 1994. In both those years there were discussion panels at Microprocessor Forum on the future of microprocessor design where some panel members (Prof. John Hennessy from Stanford and co-founder of MIPS and Tom Blank from Maspar at MPF '93 and Deutsch at MPF '94) predicted that on-die SMP would be the integration strategy for the future. It now looks like the advocates of on-die SMP were almost a decade ahead of their time. In the intervening decade, designers have eeked performance out of longer pipelines, bigger caches, and VLIW. Having now exhausted those techniques, we're back to on-die SMP.

Who Has Benefited From This Competition?

Ignoring the lawyers, during all these rounds between two fiercely competitive companies, the consumer has been the overall winner. Intel pushed hard to pass and stay ahead of AMD in the late 1980s and '90s, and AMD has succeeded in keeping PC processors more affordable. Intel developed and introduced key system technologies, such as PCI, PCI Express, USB, and AGP. AMD did follow Intel on technology roadmaps until it blazed a different trail, starting with the 3DNow SIMD extensions and extending to the x64 (AMD64) instruction-set extensions and the HyperTransport interface. This new trail was in no small part owing to the merger into AMD of the more independent minded NexGen group. Consumers and OEMs did lose the ability to buy one motherboard and to choose from a variety of CPUs to populate the boards. With each CPU vendor on its own infrastructure, however, motherboard validation and reliability have improved.

What's next? Well, Round 10 will likely be quad cores. While the Intel *Ts technologies—such as manageability, virtualization, and network stack acceleration—are important to business, I don't believe these issues can be quickly and clearly articulated to the media and public. A good race needs a clear and easily defined goal, and the core number is the next "GHz" measure.

Spring IDF's Message

After the humbling Intel went through in 2004, the company is well along in its comeback plans for 2005. Of course, we're talking not about the company's very healthy financials but rather about its bruised ego. The message I took away from the most recent Intel Developer Forum was that Intel has fully embraced its new processor roadmap strategy of dual-core processors and is ready to execute that strategy. Demos of dual-core processors abounded; even the 65nm Yonah processor, which is a year away from introduction, was shown in the demo pavilion. Intel wanted to make sure everyone would have high confidence that the company would execute on its plans in 2005 and into 2006.

The message at IDF is very clear: dual-core/multicore processors are coming—ready or not. We're just waiting for the ding of the bell to start Round 9.