Wednesday, October 18, 2006

Intel's dead ends

I didn't make emphasis on Intel's products going forward in my latest post.

Current Conroes are good products, superior to what AMD is offering. Intel should make use of that while that situation lasts. Is Intel doing it?: No. Intel is trying to prevent AMD from gaining more market share sacrificing the premiums it may charge from Conroe. The net result is counterproductive of this strategy: On one hand Intel is selling cheap products that it hasn't been able to get to the best binsplits and production volumes; and on the other AMD is nimble enough to change the production mix to mid range dual cores rendering the attack ineffective.

But with AMD entrenched deeper in the mainstream segment, PentiumDs are displaced further low, pushing Pentium4s that push Celerons, creating a catastrophic situation for Intel expressed in the growing himalayan mountain of inventory.

This situation also denies AMD the chance to market/produce Semprons (although Semprons are 64 bits), but this isn't much of a problem if it is due to higher ASPs due to a shift to its dual cores.

Intel's suicidal strategy

Thus, Conroe's aggressive pricing seems to be having only the effect of depressing the whole Intel market to merely trim some high end and low end AMD profits.

On the other hand, Woodcrest is acknowledged by the industry as a dead end. That's why the 90nm, old design of the Opterons Rev. F stole its thunder, because Rev. F is replaceable by true quadcores that may more than double the performance of the computers based on them; while Woodcrest in a childish "me too" offers Clovertown, that is scoffed at by the market.

The problem is again the front side bus. Intel can only scale up in the number of cores by increasing already huge intra-core caches, or partitioning the memory space in several buses that must have extremely complex and performance zapping logic to maintain consistency among them. It is the problem of the contention of the bus that gets worse exponentially on the number of cores.

This leads me to something very important: Intel's apparent lead in multicores is actually a suicidal strategy, let me explain why: Intel doesn't have any multiprocessor interconnect system, thus it can not even do decent dual processors [note: Woodcrest works in dual processor configurations, but it requires complex logic and the performance gains are not good], thus it can only avoid the FSB contention problem by integrating and sharing the caches among the cores, the way it is done in Yonah, Conroe, Merom, and Woodcrest. But Intel doesn't seem capable of doing a true quadcore, that is, a processor that has integrated in the same die all four cores. If it can not do that, but only the Multi Chip Modules (MCMs), then one half of a processor is forced to communicate with the other half through the Front Side Bus, just like it happens with the Pentium Ds. That is why Kentsfield (double Conroe) and Clovertown (double Woodcrest) will eventually be a failure. Especially Clovertown: A dual Woodcrest computer, that already has the FSB saturated with four cores competing for access, won't be able to keep up with eight cores.

But Intel insists in quadcores. That is excellent news for AMD, because Intel is willing to be the pioneer that receives all the arrows in the back to legitimize the market for quadcores, while AMD is preparing a solution superior by leaps and bounds that will have the field well plowed for its introduction. Once AMD quadcores are available, all doubts about Intel's inefficacies will be dissipated.

In a more abstract way, Intel is only playing to key AMD's ccHTT strenghts by insisting in the multicore route, and that is strategically suicidal.

About Itanium, I will only say that it sadistically pleases me to see that Itanium is gaining a tiny bit of market share from Sparc and others as to delay the "cut the losses" decision that Intel must eventually take. I mean, it is clear that Itanium is going nowhere but the grave, it is good to see that Intel finds reasons to keep distracting huge resources into a futile endeavor.

Then, we must speak about the 45nm. So far, AMD's 90nm products have been very competitive with Intel's 65nm, it shouldn't be like that at all, 65nm should be killing cold any 90nm products; perhaps AMD is having very advanced transistors that are 65nm in everything but feature size. What we know is that today's AMD's SOI at 90nm is as power efficient as Intel's Strained Silion at 65nm. Since Intel wants to get ahead of AMD somehow they may be going in directions that are strategically inconvenient. Perhaps Intel is again playing to AMD key advangates by running as fast as they can to the 45nm, because all the problems of power leakage get worsened the smaller the feature scale, and that highlights even further the advantages of Silicon On Insulator. Being as it is that AMD is preparing toghether with IBM 45nm Strained SOI technology, and is depeveloping the smashing technology of ZRam (ultra cheap caches that make use of SOI effects); it may very well be closing the gap in silicon processing.

Intel said that during this quarter they crossed over to 65nm, meaning that they are now (only now!) producing more 65nm products than 90nm, no wonder AMD is doing so good.

In conclusion, it seems that Intel is acting reactively just the way AMD wants it to act, and looking forward in Intel's strategic lines, we see AMD all over the place.


Anonymous said...

Intel is playing old cards that don't work anymore; because AMD has a good brand and it is higher end.