AMD is at the crossroads

AMD'S LATELY been in the news quite a bit - mostly for the wrong reasons.

Repeated problems and embarrassments with the upcoming quad-core CPUs, from new bug problems to frequency disappointments, not to mention the aftermath of ATI (in)digestion. And all that while some of the bosses enjoyed some extra $$$ in their pockets too.

And, all that at a time that can't be worse for the company. Intel is now at its strongest in the X86 performance field for the past five years. Financially it's able to afford as many depreciable fabs as it wishes, and any yield problems that could happen. Even if it was, in theory, to sink the Itanic ship tomorrow, all that would happen would be a slight rise in its share price.

You've seen Charlie's recent story about desktop Phenom being pushed to early next year and, even if AMD has a few dual-socket QuadFX-type Phenoms, it might only really ship those at Christmas time. At the same time, new missives to the OEMs are about Barcelonas - technically the same things as desktop Phenom, but with more HT channels and slower memory controller at DDR2-667R vs DDR2-1066 - reaching them in August.

Why such a large gap there? Something's just not right here - unless that memory controller could be a problem in desktop units, but not in slower server chips.

I was also quite mystified by those Povray Barcelona benchmarks circling the Web. AMD would have surely been aware that, especially if half the FP is disabled, the FP-intensive Povray would suffer. But the choice of a benchmarks which performs and scales beautifully on Core 2 - I got 7.8x scaling out of 8-core ClovertownDP by default - perplexes me - AMD would have known in advance that any leaked bad results can only be detrimental. Dailytech's recent quickie Cinema4D bench run - before they were asked to leave by AMD - proved the point: clock-for-clock, Barca was slower than the current Clovertown.

Then came the talk of the fab selling. Generic processes without tuning are an enemy of top-notch CPU design - the final CPU speed, power draw depend so much on tuning and squeezing the best out of the process, therefore controlling the process from A to Z. Remember the Alpha? Besides its straightforward architecture, it depended on the highly tuned DEC Semiconductor process and design by hand to get its astounding speeds. In a generic fab, it couldn't do much - Samsung took quite some time mustering the DEC process tricks to fab Alphas in 1998.

Fabless CPU brands, especially the X86 ones, don't exactly catch the high performance marks, as large fabs wouldn't tune their whole process for a specific CPU run - unless a correspondingly large wad of cash is thrown on the table. If that's the case, you may as well keep the fab.

Once you lose control of the fab, the new owner ultimately has the say what it will focus on, and if something else gives better returns than the AMD CPU run, so be it. So, even if in a financial quagmire, AMD really shouldn't give up the ownership of any of its up-to-date fabs.

Unless, well, a remote possibility happens of that new partner being an extremely friendly party with a common goal - challenging Intel for overall CPU market domination in the long run. Such party would want to have AMD's fabs, design teams, X86 across-the-board market presence and, why not, combined CPU and GPU offerings, even in the Fusion combo. Even better if that party may have some hidden old grudges against Intel, since the time the first PC came about.

Did I say IBM? To me, it presents the most logical "salvage partner" to AMD, and the only plausible fab co-ownership choice. It was among the first Tier 1 vendors to step in supporting Opteron, it has a strong money-making semiconductor unit - all three consoles depend on IBM for their CPUs - and they do want Power to continue its run against Itanic at the high end. Combine it with the current semicon process collaboration and proposed Power7 - Opteron HT3++ interconnect and socket compatibility, and you got a clear answer.

Some kind of "JV", I can't really call it anything else, combining these AMD and IBM semiconductor resources, including top-notch fabs, would make sense. And, if properly and strategically run, it would present Intel with a pretty much unified competitor spanning the whole range, from high end Power and Opteron CPUs sharing a common platform one day, to mainstream X86 desktop and mobile chips, chipsets and GPUs - both the graphics and GPGPU kind. And, yes, the only serious semiconductor process competition to Intel, too.

Also, the stronger combined (or multiple) teams and added resources would help resolve the nasty problems like the recent Barcelona stepping's new bugs or rumoured speedpath problems. The IBM gang may even have better use for those ATI GPGPUs, for instance using a variant of low-power mobile R700 with an added CPU and interconnect one day to replace the largely ineffective, tough to program, Blue Gene cores in their custom supercomputers.

The only condition for all this? New strategy and vision, i.e. new management, I guess...

The alternative? Well, AMD could find a quick outside investment fix, or even be taken private. The fresh inflow of funds would keep the fabs, while buffering the company for a couple of years to accelerate the product development, update its processes on time and, in general, get its act together.

Getting that act together might still involve management changes. Many were complaining about the effects on AMD that the change from Jerry Sanders to Hector Ruiz brought - without pointing fingers, all I can mention is the Motorola experience. Remember how, once long ago, the Moto 68K family was the sovereign rulers of the CPU perfomance space, with the first 32 bit CPU, the 68020, in 1985, and an architecture vastly superior and elegant compared to the X86? An architecture which, despite the dependence on Macs and high-end systems only, could have been continued and developed well? The wrong strategic decisions that ended it, and the later move to PowerPC, where Moto again took the shorter end of the stick?

I just hope that the Motorola cadre at the helm of AMD, Hector Ruiz, will not "Motorolise" the AMD CPU business and repeat that awful undeserving fate of the 68K. We all stand to lose terribly if AMD suffers a major setback - if nothing else, there will be far less to write about, and no one wants the return of old, domineering "Darth Vader" Intel.

Alternatively, can Jerry Sanders come back? I don't know, maybe at his age, he may find it more entertaining to help run some vastly more profitable enterprises with less technology risks or development pressures, like maybe Kentucky Fried Chicken.

No, he should come back. Because, if there's no solution in sight, the new quad core continues slipping while Intel Penryn accelerate forward, and ATI side doesn't get on a par with Nvidia in the GPU race - it is AMD, and its share value, that may end up "deep fried". µ