A Great Week! . . .

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

AMD hasn’t had a good week or so.

One, they released 2X2 cores that lost practically every meaningful test to Intel’s Kentsfield with almost twice the power consumption. Everyone was so impressed that AMD cut the price before charging it.

Two, they “showed off” a 65nm 4-core running Task Manager rather than tasks. Guess it’s better to be a manager than a worker.

Third, they “released” 65nm chips, and showed them off to nobody, and nobody seemed to mind.

Fourth, they found the need to “instant rebate” most of their dual-core processors about 10%.

Finally, it doesn’t look like the 65nm CPUs AMD plans to release about a year from now will reach the raw GHz speeds Intel’s current dual- and quad-core chips. We’re seeing 2.5GHz max for quads, 2.9GHz for duallies. Given that AMD’s current chips do rather less per clock cycle than Intel’s chips, it’s hard to see how these chips can be tweaked to exceed what Intel has today, much less the 3.5GHz/4GHz 45nm quad/duallies from Intel that will begin to show up around the same time.

(Personally, I suspect 3.5/4 is a bit too high, but even if you take 10% off the numbers, that still leaves something like a 40% performance gap between frequency and IPC. That’s an awful lot of tweaking.)

I’m sorry, but this doesn’t look good, doesn’t look good at all, especially right at the time when AMD will need to take away much more marketshare from Intel than they have up to now.

Here is a very revealing graph (which comes from the Economist (based on data from Mercury Research):


Obviously Intel has gone down a bit, and AMD has gone up a bit, but not a whole lot, despite all the real problems Intel has had and the real advantages AMD did have along with all the media stories about AMD whipping Bluehide.

Outside of the multicore, Hypertransport-critical server environment, what’s going to happen next year when AMD has more capacity and less capable processors?

Yes, the person buying a Dell or HP won’t generally be looking for the fastest thing around, but what kind of prices will AMD be able to get if Intel charges $150 or $180 for a CPU that is, say, 30% faster than its base model?

What’s The Problem?

I suspect the problem boils down to three letters: SOI, both in a direct and indirect way.

The direct part is simple enough to see, SOI is tricky stuff, and AMD has had a lot of problems getting it to go fast in 130, 90 and now 65nm.

But I don’t think that’s even the bigger part of the problem now. I think the biggest part of the problem with SOI now is that it was considered to be the solution to the power problem, and that’s proving to be not enough.

AMD initially adopted SOI for the simple reason that they couldn’t have built Hammer as it currently is without it. Given its design, a Hammer without SOI would have been a green Prescott, or worse.

In contrast, Intel’s response to the PIII crashing and burning was to try to duck the problem with architectural changes (lengthening the pipeline), then ad hoc design changes to deal with this phenomenon of transistors leaking more and more power as the speeds went up and the circuits became narrower and narrower.

Eventually, all the quick fixes weren’t enough, and eventually Intel got stopped cold, or rather, hot. A quick fix simply didn’t exist for them. So they pulled out a contingency plan (converting their notebook chips into desktop chips), plus they bit the bullet and dealt with the power leakage problem that had plagued them.

It’s the second item that’s responsible for much of C2Ds improved performance, and the rest will come with 45nm. In contrast, AMD came up with a half-solution that carried them for a couple generations, but now that’s not good enough for multimulticores and strict power requirements, and now they’re paying the price for it.

I further suspect that until AMD does a CPU redesign that is the equivalent of what Intel did, a new architecture, new physics, new manufacturing chemistry, they will continue to trail Intel in raw performance.

Maybe architectural experiments can make up for that shortcoming, but you have to wonder whether it’s a matters of making lemonade out of lemons.



Leave a Reply