Columnists

AMD: Praying for a triple, hoping for a home run

With a spate of recent bad news and a big press event slated for Monday, AMD needs to hit one out of the park. The company's very survival may be in question, so a single or a double just won't do.

By Brian Dipert, Senior Technical Editor -- EDN, 9/6/2007

AMD's got a major press event scheduled for Monday night at the Presidio in San Francisco. Widespread industry scuttlebutt predicts that the company will formally (and finally) unveil the initial "Barcelona" workstation and server variant of the K10-microarchitecture, quad-core CPU family, details of which have leaked into the public consciousness over the past four years.

Because of a news embargo I agreed to honor, I can't yet tell you what the company will reveal Monday night (though I can tell you that I plan to be there). However, regardless of when AMD finally launches Barcelona, its subsequent success (or failure) will have the most pivotal effect on the company's fortunes since the unveiling of the obtained-through-acquisition K6.

I'll explain this opinion by means of an example. The Quad FX system I used in a May 10 hands-on cover story (see "Interface overkill? Is eSATA necessary for your next system design?") was based on two dual-core, 3-GHz, AMD Athlon FX-74 CPUs. Early testers found that the chips self-identified as Opterons, which isn't terribly surprising because they shared the Opteron's 1207-pin Socket F packaging (although the die had been slightly altered, perhaps via a single- or few-metal mask tweak, fusible link reconfiguration, or other method, so that the FX-74 could use non-ECC versions of DDR2 SDRAM).

AMD sells the FX-74 and its FX-7x siblings only in two-CPU bundles, reminiscent of AMD's Opteron 200 series CPUs. However in late February, AMD unveiled a Socket AM2-based, 3-GHz, dual-core CPU called the Athlon 64 X2 6000+. Compare its specifications with those of its FX-74 equivalent, and their common Opteron lineage becomes clear. By eliminating the inter-CPU link, AMD was able to shoehorn the die into the lower-I/O 940-pin Socket AM2 footprint. In fact, Opteron 100-series CPUs, which also come in a Socket 940 pinout, are effectively identical to their Athlon 64 and Athlon 64 X2 peers. This synergy was particularly evident prior to the Athlon 64 X2's embrace of DDR2 memory, coincident with the move from Socket 939 to Socket AM2.

It isn't just AMD's desktop, workstation, and server CPUs that leverage a common silicon design across multiple product offerings. As I've also written before (see "Turion 64 X2: A Credible Choice?"), the company's Turion 64 product line is essentially a relabeling of Athlon 64 equivalents, underclocked and run at lower voltage to save power at the expense of speed. Compare AMD's approach with that of Intel, which crafted the power-optimized Pentium M microarchitecture solely for mobile-computing applications and subsequently extended those innovations into all computing segments with the follow-on Core microarchitecture.

Ironically, AMD's next-generation mobile CPU family, "Griffin," won't be K10-based. Instead, AMD plans to litho-shrink today's Turion 64 X2 to the 65-nm lithography process, which will enable the company to beef up the amount of on-die cache and make other modest design tweaks, including multiple on-die voltage planes, for incremental potential power savings. Inevitably, however, a K10-based mobile CPU will eventually emerge—assuming AMD survives long enough to design and manufacture such a product.

Admittedly, things are looking pretty grim for AMD at the moment. Barcelona is extremely late, thereby making the "Phenom" follow-on derivative for consumer desktop PCs extremely late, too. And judging from the initial 1.9- and 2-GHz launch speeds, Barcelona is quite slow compared with its competition. The company has historically touted its higher processing efficiency than Intel at a given clock-speed metric, but it's now competing not against the underwhelming Intel NetBurst microarchitecture but instead against the far superior Core follow-on.

Why is Barcelona so slow? Theories abound. Some folks claim that AMD is struggling with lingering leakage-current issues on its 65-nm process (a problem that Intel knows well from its 90-nm-lithography history). Others believe that a fundamental chip-design flaw is to blame.

AMD showcased two 3-GHz, quad-core-based systems at a late-July media day, in an attempt to dispel doubts about the performance potential of the Barcelona design. I was encouraged to see that these processors didn't rely on exotic cooling and generally seemed to be thermally (and otherwise) stable. However, producing two cherry-picked CPUs is a far cry from consistently churning out millions of units. If AMD could make 3-GHz processors in high volume, it certainly wouldn't be limiting its public declarations to claims of a CPU 33% slower than that.

Ironically, AMD's been producing 65-nm-based Athlon 64 X2s since late last year, but again only at low speeds. It remains unclear whether this deficiency is due to process problems or because the chips' redesigned, longer-latency L2 cache configurations would make them performance-uncompetitive with their 90-nm-based peers at leading-edge clock frequencies.

Arguably, AMD's past success had as much to do with Intel's stumbles as it had to do with the inherent strengths of the K6 and follow-on K8 microarchitectures (and products based on them). Intel's much-publicized setbacks included:

• Acquisitions and corporate-focus dilution into other application areas, such as wired and wireless communications, consumer end products, and ARM-based CPUs and support chips.
• The Itanium microprocessor, and a conscious decision to "hold back" the Pentium series in order to avoid internally competing with the self-anointed Itanium successor.

Today's Intel is a far different—and far more focused and formidable—competitor (with unclear legal consequence). As AMD struggles to ramp its 65-nm process, Intel's already moved its entire x86 CPU product line to 65 nm, with 45-nm products waiting in the wings. And whereas Intel doesn't yet offer a quad-core, single-die product, few real-life applications have shown Intel's dual-die, shared front-side-bus approach to be the Achilles' Heel AMD claims it is.

AMD has retained a modicum of market share by virtue of its aggressive price cutting, therefore keeping it competitive with Intel in performance-per-dollar and power-per-dollar terms, even though it can no longer definitively lead on an absolute performance or power-consumption basis. But I think you'll agree with me that short-term slashing does not a long-term sustainable strategy make.

Recent AMD news paints a pretty bleak picture:

• Severe cost-cutting, including layoffs, a hiring freeze, land sales and rumored departure from the fabrication business (who was it that said "Real men own fabs"?)
• Deceptive benchmarking data
• An executive vice president-via-acquisition resignation in early July
• Restructured debt
• The announced resignation of the company's chief sales and marketing officer two weeks ago
• Followed by the resignation of the company's senior vice president of worldwide sales earlier this week

Can AMD recover from this seeming tailspin? Will U.S. and international courts assist it in this mission by siding with it against Intel, and if so will they act in time? Nearer-term, what will AMD reveal next Monday night? Stay tuned for my follow-on report.

Note: Apologies for the baseball analogy in the headline, but the stretch drive to the playoffs is underway, and my beloved Cubs are tied for first place as I type these words late on a Wednesday.