The new AMD roadmap restores a bit of confidence

7 May 2008 | 20:14 BST

By Charlie Demerjian

AMD HAS TAKEN the first baby steps in fixing its credibility problem with the server folk, aka coming clean. It was the same thing Intel did during the dark days of the P4 era, restoring confidence in the roadmap by shutting up and doing.

Today's revelation is an admission of failure on a massive scale, both in terms of current competitiveness and future cores, but it is better than carrying on like nothing is wrong. The two biggest things are they aren't up to snuff core for core, and Bulldozer is looking wounded if not dead.

The first part is obvious K10h/Barcelona is not competitive with Penryn much less the big bang that is Nehalem. To counter this, AMD is upping the standard core count by 50 per cent to six, basically make up for deficiencies in IPC by upping the raw number of cores you throw at a problem. Die area to make up for design, like adding cache to a P4. It works, but is far from elegant.

The good news? It will make up for a lot, and the firms compartmentalised architecture will make it a fairly easy thing to do. The bad news? Intel has six-core parts coming as well, and Nehalem is compartmentalised so they can.... do the maths.

The disappearance of Montreal was at first troubling, but the addition of Istanbul makes up for a lot of it, what you lose with two cores you make up for with no additional HT hop latency. Well, not really, but it will help.

If Sao Paolo and Magny-Cours are earlier in 2010 rather than later, they might make a difference. If they are late, we really have to question why they would bother. These parts will be up against Westmere on 32nm, a much smaller, cooler and core for core, vastly superior parts. They damn well better be on time and exceed all expectations.

The most distressing part is Bulldozer. Much was made of it not being on the roadmaps last December, but that was a tempest in a tea kettle. The real problems have come out since then, and the only way to say it is that the architecture is a huge flop.

No, this isn't to say that it won't be fast or meet every spec that it was meant to, simply that it can't be done on a 45nm process. This means that it will be on the 32nm node pushing it out to late late 2010 best case. It is probably a good time to mention that restoring confidence in the AMD process roadmap might be a good thing right about now.

In essence, what AMD has now is what AMD has until 2010 or 2011. It doesn't win many benchmarks now, and is facing the biggest leap in Intel CPU power since the Core Number Numeral transition at the end of 2008. They are going to trade raw area to cover deficiencies, and it should work OK but won't win many converts.

Intel on the other hand doesn't really have to do much at this point. They can simply coast for two more years and maintain an easy lead. My guess is that they will leave the low end to AMD and reap the high-end profits while hopefully not becoming fat and lazy.

Much as many would like to point AMDs current situation at external sources, mainly Intel, they can't. AMD simply botched its core business and can't seem to get a CPU out the door at promised frequencies to save its life. And that is its own fault.

With the Barcelona debacle mainly behind it now, Shanghai looks to be a baby step forward, a quiet manageable baby step. There are tons of samples floating on the part right now, and it looks like a decent move forward in performance.

If AMD transitions this into bigger and bigger steps, it will restore confidence, have much more competitive parts, and probably survive. If they screw up again, I don't think that even Intel sandbagging will save them. AMD has laid out an adequate and more importantly doable roadmap. Now they have to do it or die. µ

© 2007 Incisive Media Investments Ltd. 2007