The Color Of Silicon

Competition in sports is a simple zero-sum spectacle; one winner, one loser, and the audience goes home excited but otherwise un-enriched by the outcome. Competition in the business arena is more complex. Everyone gains something, no matter who gets bloodied. Witness the never-ending struggle between Intel and AMD.

The two firms have been going at each other since 1991, when AMD rolled out its AM386 clone of Intel's 80386-DX. Some readers may recall that Intel licensed its 386 technology to AMD, enabling Intel's first and now its only serious competitor in the PC microprocessor market. Far from hurting Intel's sales, AMD has improved its archrival's products and revenues while thriving in its own right. To explain this ironic result, let me turn to a classic movie about my favorite subject, billiards.

In "The Color of Money," Fast Eddie Felson (Paul Newman) is a washed-up hustler with deteriorating vision and little motivation to earn more than his daily bread and booze. One day, he is awakened by the "thunderbolt break" of Vince Lauria (Tom Cruise), a talented and brash player who comes out of nowhere. Vince's meteoric success inspires Felson to get serious about his own game again. Inevitably, Felson and Vince meet in a winner-takes-all tournament. There is no decisive victory; and many movie fans are disappointed by the apparent anticlimax. But they miss the point, just like those who seek a definitive winner in the Intel v. AMD match.

Intel without AMD would be like Felson sans Vince: uninspired, inert, and sloppy. We might all be using 80486 machines today. Conversely, AMD's game would not be so sharp if not for the challenge posed by Intel's preeminence. Indeed, AMD might not even be in the game had Intel not licensed its technology to AMD way back when. Like Felson and Vince, Intel and AMD bring out the best in each other whenever they go head to head for serious money. The coming year's silicon tournament features two main events: dual-core and 64-bit x86 processors.

The Race To 64 BitsIntel broke the 64-bit rack with Itanium back in 2001. But it was a weak break that left Intel behind the 8-ball. Itanium's IA-64 instruction set ran 32-bit applications slower than comparable x86-based CPUs, and few customers were interested in rewriting all of their code to take full advantage of IA-64. Even the much-improved Itanium 2 is limited to a small universe of very high-end server applications. The richest action lies in x86-compatible workstations and servers.

AMD introduced its Opteron processor with 64-bit x86 extensions in April 2003. Not until July 2004, did Intel match Opteron with a 64-bit x86 Xeon processor (formerly code-named Nocona). Did Intel lose that match or intentionally sit it out because the big money had yet to arrive?

A 64-bit processor without a 64-bit operating system isn't much use. Microsoft was supposed to release 64-bit versions of Windows XP and Server 2003 by the end of 2004. Intel's timing of Nocona seemed ideal. But just weeks later, Microsoft again postponed its 64-bit ship date, this time until mid-2005. Linux is still the only operating system that fully supports 64-bit processing with x86-based CPUs. While Intel sat on the sidelines awaiting a ship that hasn't come in, eager contender AMD honed its skills and earned respect among the other Vinces of the tech world.

"When (Opteron) was in early design, AMD provided simulators as well as contributing code to the Linux, GCC (the GNU Compiler Collection), and other open source projects, which engendered good will among the developers," says Jason Perlow, president of systems integrator and consultancy Argonaut Systems in Tenafly, NJ. "The 64-bit enhancements in those open source projects are commensurate with (AMD's) contributions and willingness to work with the OSS developers.

"On the other hand, the (Itanium) IA-64 Linux kernel was developed almost entirely in-house at Intel and HP, and the EPIC instruction set improvements in GCC are practically nil. So really, 64-bit code in Linux on Itanium is totally un-optimized." Argonaut recently procured 500 Opteron servers for a Linux cluster in a digital effects studio. "Itanium I never recommend," says Perlow, citing the need for expensive custom programming and proprietary compilers.Nocona, representing the new improved Fast Eddie Intel, is better positioned to compete in the mainstream server market. But is Nocona good enough to beat Opteron? Not yet, if one believes early benchmark tests conducted by hardware analysts at AnandTech.

The tests, published August 13, matched AMD's Opteron 150 against Intel's Nocona Xeon running SuSE 9.1 Professional 64-bit Linux in default and smp modes. The Linux configuration took full advantage of the Xeon's HyperThreading mode. Even so, Opteron blew the doors off Xeon in database, encryption, and rendering benchmarks. Opteron's lead shrank in chess and Gzip compression benchmarks. It should be noted that both processors are workstation products. The server match has yet to come.

Opteron crushed Nocona in an Apache Web server test conducted by GamePC, a builder of custom high-performance PCs. The test used Apache's internal benchmark to simulate a 100,000 client load. The 3.4 Ghz Nocona processor lagged far behind even a 2.0 Ghz Opteron 250 in requests processed per second and peak transfer rate, performing only 75 percent as well as its "slower" rival. Nocona hit only 64 percent of the 2.4 Ghz Opteron's numbers.

Intel miscued on Nocona, but it's too soon to tell how badly. Remember that Intel's eye is on the money ball - Windows. The 64-bit beta versions of Windows were tuned for Opteron, naturally, since that was the only x86 platform available until July. Microsoft will tweak Windows64 for both processors by the time you read this, but it will be a while before we see Fast Eddie and Vince shoot it out on a level table. Furthermore, 64-bit x86 benchmark tests are still works in progress. We can't yet be sure of either Opteron's or Nocona's performance in absolute terms. Still, there's something fundamentally different about the players' techniques that gives Opteron an edge in server applications.

AMD's Direct Connect architecture directly connects the memory controller, the I/O, and additional CPUs to the central processor. Even Intel's fastest 1033 Mhz front-side bus poses a crippling bottleneck to throughput. AMD's HyperTransport bus technology gives it another edge, enabling multi-port data transfers at up to 19.2 GBytes per second. (Yes, Bytes, not bits.) So what's Intel to do? When you can't win with your own cue stick, borrow your opponent's.Intel has already adopted AMD's AMD64 64-bit extensions to x86; with trivial variations, Intel calls its version Extended Memory 64 Technology, or EM64T. Nothing prevents Intel from adopting the royalty-free HyperTransport technology. That leaves only the front-side bus problem, which Intel will presumably fix before its next major match with AMD in the dual-core processor tournament.

Single-core processors have hit an insurmountable wall. As clock frequencies rise, so do power consumption and heat. More power means much more current leakage (wasted watts) when one shrinks chip features below 100nm, as Intel learned with its first 90nm offerings. Even more power is required to dissipate all that heat. Power and heat inefficiencies are rising much faster than clock speed gains in single-core processors. It's time for a dramatically new approach to performance enhancement: multi-core processors.

Two 1 Ghz processors consume less power and generate less heat than a single 2 Ghz processor, thanks to the exponential relationship between clock frequency and power/heat inefficiencies. Put two processor cores on a single die and electricity bills plummet. Dual-core processors can also take full advantage of parallel processing techniques like Intel's HyperThreading.

Intel and AMD are both committed to producing dual-core server products by mid-2005. How well each firm will execute its dual-core strategy remains to be seen, but it appears that AMD will have an easier time of it than Intel.

Opteron was designed from the start with dual-core in mind. Direct Connect Architecture and the HyperTransport bus technology make adding another core to a die relatively easy. Intel, however, will have to part with its past before it can embrace the future.Intel's Xeon and Pentium 4 processors are based on the company's hoary Netburst architecture, developed in the days before power and heat became critical problems. The newer Banias/Centrino architecture of the Pentium M consumes much less power than Netburst, and two Pentium M cores would be easier to integrate on one die than two power-hungry Netburst cores. Yet rumor has it that Intel's first dual-core processors will use Netburst. Ironically, Intel's reputation as the industry standard handcuffs it to this antiquated architecture.

Banias is roughly as young as AMD's Opteron architecture. But AMD customers are used to gambling on new innovations, while Intel buyers value tried-and-true platforms more than breakthrough advances. Intel server buyers do not want to absorb the shock of dual-core and Banias simultaneously. So Intel is virtually forced to enter the dual-core market at a disadvantage. Bet on AMD to beat Intel in the first match of the dual-core tournament, at least. But Intel will come back from behind, as it always does.

When the last ball is pocketed in the dual-core tournament next year, Intel and AMD will be tied for second place behind their customers. Both players will provide solid 64-bit, dual-core processors that take server performance to dizzying new heights. Then the price wars will begin again, making server buyers even bigger winners.

Second place will pay off handsomely for Intel and AMD. The combination of 64-bit and dual-core technologies will bring billions of dollars to both firms and their partners, as buyers rush to replace those strained, overheated, power-guzzling processors. But don't expect Intel and AMD's market shares to change much. Buyer behavior has enormous inertia, and server purchasing decisions depend on much more than the differences between CPUs.

"As a server buyer, I'm more concerned about the servers that IBM, HP, and/or Sun will have available to support the architecture," says Robert Nocera, CTO and partner in NEOS LLC in Manchester, CT. "I've always looked more at the server as a whole than which chip runs it. A great chip will only get you so far if the rest of the architecture can't keep up. As a company, NEOS has used whichever chip can give us the lowest price for the performance we need and we've never really experienced any technical problems with either" Intel or AMD. Nocera and NEOS represent the vast majority of rational buyers. Then there are the two extremes of the buying spectrum.Some people will always buy Intel because it's the safe thing to do. Others will cleave to AMD because they love an underdog just like themselves, especially a scrappy innovative one like AMD. Any significant change in either chipmaker's market share will be temporary, eliminated as the two players adjust their pricing and marketing expenditures. Neither company can grow significantly at the other's expense.

The only way to increase the pot in a pool tournament is to draw bigger crowds. Better play is what people will pay to see. The competition between Intel and AMD is what keeps both firms' revenue growing, and the computing industry moving forward. Neither firm can afford to win decisively.

David Hakala is a freelance technology writer based in Colorado.