To Be 3, Or Not To Be 3? . . .

Some aren’t too happy that I didn’t respond properly to AMD’s 3GHz demo, and told me so.

I think reproducing an email exchange I’m having will illustrate the issues pretty well:

Why The Skepticism?

Q. Nobody is pleased about where AMD but lets not get ridiculous. First, you are assuming the K10 sample is the “sweetest, most-blessed engineering sample”. How do you know this? Maybe they finally got it right and we will see the higher speed bins at the end of this year and early next as AMD has promised. Couldn’t this be logical as well?

A. No, because they haven’t budged on the anticipated speeds of Phenoms, see .

First, giving ranges of speeds for specific processors ought to tell you something about the situation. The low end of the range is probably what AMD can deliver now. The higher-end is what they’re hoping really, really hard they might be able to do some months from now. Neither gets close to 3GHz at the X4 level.

Followup Isn’t it better to over under promise and then over deliver rather than vice versa? Why are you so skeptical of their 3 Ghz sample? When it was a 1.8 ghz sample at computex we all took it for what it was. Do you think they were piping liquid nitrogen into the box at the analyst meeting from below the table? Dosen’t it make sense form a manufacturing perspective that if they were at 1.8 during computex that the processors that they launch in Aug will be running at around that speed? So if they are able to get samples now at 3ghz why would it be any less surprising that they would be able to easily attain their 2.6 Ghz roadmap figure with headroom to spare?

A. I’m not skeptical about AMD being able to make a handful of 3GHz, or that the demo ran at 3GHz. I am quite skeptical that they could make enough to sell you one now or anytime soon. No, I didn’t think or say at any time they used LN2, the pictures indicated that obviously wasn’t the case. All I said was that it was likely a cherry-picked CPU, which now seems to be the case.

Yes, the article says that “the parts were sorted, but not heavily,” but, to give one possibility, if you did a wafer with a formula shown to provide a low yield of high-speed parts, you wouldn’t have to do too much sorting.

Some overvolting of the chip seems likely, though probably not too much (strange that the person who claimed he saw 3GHz in the BIOS didn’t try to find out the BIOS reading of the voltage and temperature).

We don’t know exactly what AMD did to get that CPU running at 3GHz, and in the great scheme of things, it doesn’t matter. What matters is AMD being able to sell the public a quad-core that can run at 3GHz. So far, they’re not promising anything even near that, and from what little we know, it looks like they’re going to need a 3GHz to be competitive against Penryns.

Why Do You Think They “Suck”

Q. Further, where do you come up with ‘they must suck’??? Just because you don’t have benches? You, yourself have extrapolated the performance. Do you somehow think K10s are not going scale normally? Just bc you don’t get benches is really no reason to go off the deep end with an article about them ‘sucking’.

A. But we have a couple benches, the Cinebench and the Pov-ray, both of which scale directly with frequency, and they both indicate that K10 is just a tweaked K8 and don’t look terribly competitve against high-end C2Ds, never mind Penryns. The two spec benchmarks AMD keeps universally citing may have some applicability to server use, but are pretty dubious desktop benchmarks.

Followup: Barc is a server processor after all.

The device being tested was a Phenom, but that’s a minor point. The real issue is AMD’s refusal to have its systems benchmarked.

Q. Unless AMD has flat out lied about their performance estimates 3GHz Barcs put them back in the game to match or exceed Xeons for 08. It is actually a quite positive turn of events for green that you seemingly missed.

A. No, AMD doesn’t lie, it misleads. If the spec benchmarks aren’t representative of overall real performance, and that’s all AMD puts out, that’s as untruthful as fabricating numbers. Second, while low clock speed is the major problem with K10, popping out a well-guarded demo system while not changing the real indicators of what AMD can deliver is just a stunt to fool the credulous.

That’s what angers me about AMD so much. People like yourself want to believe, want to think the best about the company, and they’re taking advantage of your goodwill. They tell you stuff that sounds good until you closely examine what they actually said, and you find that what they said wasn’t quite what you thought they said. In the meantime, while folks like you are distracted, they slip in the far less pleasant details for which they’ll be held accountable at some future date.

It’s not inconceivable that something like a better SOI process from IBM might make something like a 3GHz K10 possible in six to nine months. However, it’s very clear that they don’t have it now, and just as clear that without some future big breakthrough, K10s production will be almost entirely low-speed (i.e., less than 2.5GHz, probably even less than that).

Followup 1: Wow Ed, that is a real leap there. If it is x% faster in spec than it is x% faster in spec, period. How exactly do you get from associating spec numbers to fabricating numbers?

A. Different benchmarks report different relative results. If that weren’t the case, all we would need would be one benchmark. Sometimes, a CPU can handle a benchmark relatively much better or worse than the majority of other benchmarks. While that’s quite important if you use the specific program on which the benchmark is based; it’s a bad indicator upon which to based general comparisons. This is why you need an array of benchmarks, to sort out the typical from the untypical.

Let’s look for a moment at the Cinebench and Pov-Ray benchmarks. When used to compare current AMD/Intel processors, Cinebench is pretty representative of relative AMD/Intel performance, if anything, it favors AMD a little bit. As I’ve shown in the past, K10 does little better than current AMD processors and would have to run about 15% faster than a Penryn to match it in Cinebench.

In contrast, Pov-Ray is a negatively unrepresentative benchmark for AMD processors; they do much worse against C2Ds than the average benchmark indicates, and K10s show little improvement. What was significant about the K10 Povray benchmark was not how badly it did against Penryns, but the small improvement they showed over current AMD processors.

So not all benchmarks are created equal, but that’s not the important point here.

When asked, “Why don’t you show more benchmarks?” AMD’s answer boils down to “We don’t want to give Intel an advantage.”

Let’s think about this for a moment. If AMD truly didn’t want to do this, and these spec benchmarks truly indicate that K10 is a lot better than anything Intel is going to have, didn’t they let the cat out of the bag by releasing the spec benchmarks? Wouldn’t Intel be able to read the handwriting on the wall better than just about anyone else, and start making whatever adjustments they could?

If those spec benchmarks really meant general, overall superiority, and you told your competitor that months ago, what possible competitive harm could possibly be caused by piling on with lots of other benchmarks showing that same overall superiority? Intel certainly did when they put out C2Ds benchmarks well before product introduction.

No, AMD’s benchmark position doesn’t come from Intel, but Apple: find the benchmark in which you do best, then trumpet it as being typical.

No, what AMD is doing makes no rational sense if they have the draw on Intel. It only makes sense if they don’t have the draw on Intel, but want to fool you into thinking that they do.

So AMD isn’t actually lying when they say, “We don’t want to give Intel an advantage.” The problem is that doesn’t mean “We don’t want Intel to redesign their chip,” but rather “We don’t want Intel to spit out better benchmarks now.”

P.S. If AMD really didn’t want to give Intel a heads-up, then why would they reveal K10’s design features long before even “simulated” benchmarks? If there is anyone in the world who could extract the maximum amount of information from that, wouldn’t it be the Blue Chip Architects and Friends?

Followup 2: I don’t get it, how is it very clear that they don’t have it now? They showed it. When Intel shows their 3.3 Ghz Penryns nobody bats an eye but when AMD demos their 3 Ghz Barc there is a conspiracy theory, why?

A. I’ve already gone into the difference between showing something and being able to sell it. The major difference between AMD and Intel is that Intel has not only shown a 3.33GHz Penryn, they’ve also said they’re going to sell some shortly.

No, that doesn’t mean the top guys at Intel resemble the Holy Trinity. They played a couple games, too, like including a bunch of SSE4-enhanced benchmarks, but they also provided more typical benchmarks which showed that Penryn was basically a tweaked C2D.

Now if Intel had shown only SSE4-enhanced benchmarks, then you might make a case for moral equivalence. But they didn’t.

Intel also knows that silence is golden sometimes. They put out enough numbers to satisfy those interested, then sit back and let AMD make a spectacle of itself.

Shooting The Messenger?

There’s something very odd happening among those who write me about this: they don’t really disagree with what I’m saying. Instead, they seem to be unhappy (sometimes very unhappy) with my saying it.

Or maybe what they really don’t like is that they want me to give at least some hope of a happy ending to the story, and I haven’t exactly been giving them one.

Well, what I’m seeing is a company with not too much money losing much of it each quarter, and the products that are supposed to bail them out of the tailspin don’t look too competitive, and even if they are, it doesn’t look like they’ll be able to make a whole lot of them anytime soon.

I see a company that acts like some sugar daddy will bail them out at the snap of their fingers, but I don’t see Daddy.

I see a company that seems to be running itself into the ground, and the executives alternate between putting on an Alfred E. Neumann grin and saying, “What, Me Worry?” when they’re not acting out Little Orphan Annie singing “Tomorrow.”

What, me worry? Yes, me worry. Maybe you should worry, too.


Be the first to comment

Leave a Reply