By Brian Dipert, Senior Technical Editor -- EDN, 12/14/2007
Apologies upfront for the delayed completion of this particular writeup, coming more than three weeks after AMD unveiled its Phenom microprocessor and the bigger-picture Spider platform, which also encompasses a next-generation core-logic chipset and graphics processor. One upside to the delay, the result of a variety of (both planned and impromptu) contenting personal and professional factors, is that I've been able to amass a more encompassing analysis of Spider's (and AMD's) current health and future prognosis than I might have been able to do via more knee-jerk intro-date coverage. Unfortunately for AMD, however, the statistics and insights I've accumulated over the passage of three weeks' time have dimmed my view of that health and prognosis. Before proceeding, please peruse my colleague Suzanne Deffree's outstanding coverage from yesterday's dour financial analyst briefing as background:
AMD seeks profitability, says Barcelona botch won't be repeated
AMD: Let's start with what's in the glass …
At the moment, AMD is reeling on a number of fronts due to a number of factors...technical, financial, managerial, etc. I'll cluster my thoughts to follow in a few big-picture topic categories.
Microprocessor speed bumps
I had a hint of the trouble to come when I noticed that only 20 pages' worth of the 74-page briefing presentation I went over with AMD prior to the Spider launch were devoted to Phenom, and that AMD's pitch led with a 28-page foilset on the ATI Radeon HD 3800 GPU series. Phenom's underwhelming introductory clock speeds leave it at a competitive disadvantage when benchmarked against current Intel CPUs, a gap that only grows when Phenom is compared against Intel microprocessors to (shortly) come (one of which I have in-house).
Post-intro, we've learned that both Phenom and Barcelona currently suffer from a TLB (translation lookaside buffer) L3 cache controller-related flaw (which, predictably, AMD's biggest fanboy inappropriately plays down...one wonders how Yager's editorial would read if the flaw was Intel's?). AMD has temporarily halted Barcelona shipments as it redesigns the chip to fix the bug, with the exception of a few specific customer cases where the particular application has been judged to not expose the flaw. But mind-blowingly, AMD has decided to continue shipping Phenom, thereby suggesting desperation in the company's financial situation.
The TLB flaw is particularly probable when the system is doing O/S and application virtualization tasks, explaining the server-tuned Barcelona shipment near-halt. And AMD has also come up with a BIOS-based workaround for the bug, one that (depending on which analysis you read) clobbers performance by 5-20%, further widening the AMD-vs-Intel speed gap. But AMD's current naiveté is reminiscent of competitor Intel's in the early FDIV debacle days. How can AMD guarantee that its Phenom customers won't do virtualization or other tasks, either now or in the future, which would accentuate the flaw?
Engineering, as I've written plenty of times before, is (unlike basic science) the art of going beyond what you can do to what you should do, balancing technical feasibility against incremental cost (as compared to incremental profit), schedule impact, and other practical considerations. AMD regularly lobs sarcasm at Intel's "inelegant" dual-die, quad-core to-date approach, but to date Intel's supposed front-side-bus swamp hasn't shown itself, except in a few niche applications and with carefully crafted AMD benchmarks.
And, as any of you who (like me) have designed ICs knows, a linear increase in random logic transistors correlates to an exponential increase in design complexity (and subsequent probability of flaws). Did AMD overreach with its "true quad-core" Barcelona and Phenom and, due to subsequent schedule push-outs, enable Intel to regain momentum with the Core microarchitecture? Time will tell, but early indications suggest "yes."
A quick aside before continuing; AMD has consistently given a "no comment" response when asked about my past-published inference of substantial common design ground between Barcelona and Phenom. The TLB bug reveals just how much their lineage overlaps. With that said, Phenom supports v3 HyperTransport links (particularly attractive with the narrow-but-profitable gaming enthusiast niche), whereas Barcelona only specs HyperTransport v2 (what the silicon actually implements, albeit in a non-enabled manner, may be a different manner). AMD also claims that Phenom delivers more advanced on-chip power management hooks than does Barcelona.
Although AMD is currently shipping only 2.2- and 2.3-GHz variants of Phenom, I actually have a pre-production unlocked 2.6-GHz Phenom 9900 sample in-house, mounted on an Asus M3A32-MVP Deluxe motherboard. One snippet of the documentation I received along with the chip particularly caught my eye:
While the transistor count and most other specifications are all the same, the power and voltage ratings will be higher than the currently shipping 9500 and 9600.
AMD didn't clarify "higher," and I haven't yet fired up the board to measure it myself. But other reviewers' published feedback suggests that the incremental power draw above 2.3 GHz is substantial, substantially exponential to be exact, when mapped against a linear clock rate axis increase. I've suggested in the past, based on various news nuggets, that AMD has lingering leakage-related 65-nm process problems. This latest tidbit adds fuel to the fire. And it comes as competitor Intel is aggressively ramping its 45-nm process into full production across a multifab manufacturing network.
While reviews on Phenom have been underwhelming, those of the Radeon HD 3800 GPU series are much more upbeat. And AMD has priced the GPUs for the mid-range of the graphics card market, where per-unit profits aren't necessarily as substantial as those at the high-end (although the diminutive die sizes delivered by foundry partner TSMC's 55-nm process should help here) but demand volumes are much greater. However, the Radeon HD 3800 is only the first member of this particular product family; low- and high-end variants are necessary to complete the generation transition. Until that time, AMD must continue to rely on Radeon HD 2xxx products, which have consistently underperformed their Nvidia counterparts.
Bigger-picture, AMD (reminiscent of Ebay with Skype) has admitted that it over-valued ATI as part of the acquisition, and is going to take a write-down. I've never questioned the conceptual wisdom of AMD's acquisition of ATI; it's long been clear to me that in order to compete with Intel in the inevitable integrated graphics business, AMD would need to obtain robust in-house GPU expertise akin to the path Intel blazed with the late-1990s i740.
But from the very beginning, I shook my head with bewilderment and dismay at the magnitude of the October 2006 $5.6 billion purchase price tag. AMD also, I believe, underestimated just how substantially (as I predicted upfront) Intel CPU customers would retreat from buying ATI graphics products post-acquisition...a trend that AMD finally admitted at yesterday's financial briefing.
With this all said, don't misunderstand my stance. The computer business is at its best when robust rivalries exist in all product segments. Intel needs a healthy, pugnacious AMD to avoid becoming complacent, as has happened in the past, just as Nvidia needs a robust competitor in AMD's ATI Graphics division. And regardless of whether you're an AMD, Intel, or Nvidia customer (or source from all three suppliers), vendor complacency avoidance is ultimately in your best interests.
AMD's substantially bleeding cash reserves right now, and the timeframe of its next profit isn't clear. Can it staunch the hemorrhage and regain at least some semblance of its prior might? I sure hope so. And you should, too.