Re: Intel Inside no more


Yousuf Khan wrote:
> David Kanter wrote:
> >>>Look at SPEC CPU, the P4 was substantially faster...twink.
> >>
> >>There were a few things that P4 was faster at than P3, and that was one
> >>of them.
> >
> >
> > SPECcpu is more than just one "thing". It's a collection of many.
>
> Well, SPEC has been shown in the past to be highly manipulable by
> compiler tricks and architectures geared towards achieving high SPEC
> scores without necessarily being faster in real-life situations. That is
> unless your real life situation is running SPEC benchmarks.

Point the way to a better cross platform benchmark...

Seriously, compilers improve and SPEC takes that into account. There's
nothing wrong with that, as long as the compiler tricks can be broadly
applied. I think Sun's art hack is really dicey, but most compiler
improvements benefit classes of programs, not just benchmarks.

> I can think of several situations where SPEC has shown one architecture
> to be faster than another, but real-world applications never ran any
> faster.

So what? I can think of several situations like that for any
benchmark. That doesn't mean crap. If there is a better alternative,
why isn't the industry embracing it? There is one better alternative,
which is to use applications. However, very few people use the same
applications...therefore application benchmarks aren't quite as
meaningful to as many users.

> For example, games which are very FPU-dependent somehow seem to
> get greater benefit out of the Athlons than the Pentiums, yet Pentiums
> still somehow show up higher in SPECfpu.

That would be because it's called SPEC CPU, not SPEC video games...

> And this is not just a
> complaint that's been levelled about SPEC since the x86 processors
> started competing. It's been a complaint since the days of the RISC
> server chips competed with each other.

Yes, but you notice how everyone is still a SPEC member? That's
because it serves a very valuable purpose.

> >>However, most other things P4 was slower Mhz-for-Mhz than P3,
> >>including most other FPU stuff. As it turns out the very linear
> >>algorithms in SPECfpu benefit from higher clock speeds.
> >
> >
> > Lower IPC is not slower. I never said the P4 had higher IPC, I said it
> > was faster. That means better performance, not better per clock
> > performance. Nobody cares about the latter. Performance is what
> > matters; if you build a 10GHz CPU with low IPC, that's fine. If you
> > build a 1GHz CPU with high IPC that's fine. Ultimately, it doesn't
> > matter except for the power/heat issues.
>
> I was talking about real performance too. The P4 didn't really start
> breaking away from P3, until the P3 stopped at 1.3Ghz while the P4 was
> upto 2.0Ghz.

Hrmmm...yes, that's right. The P4 hit 2GHz shortly after it was
launched though...seems like a good product introduction then.

> Similarly, the P4 didn't really start breaking away from
> Athlon XP until XP stopped at 2.1Ghz and P4 was at 3.2Ghz.

I think Northwood probably outperformed even at 2.8GHz, but either way,
the point is that as a result, AMD got shanked and had big problems
until they could get the K8 out. That's a success. That it didn't
work at 90 was a big mistake, but the core used in prescott was not the
same as in Northwood.

>Since Intel
> couldn't really go much beyond 3.6Ghz with the P4, the Athlon 64
> completely picked it off because it had the higher IPC and a better
> thermal cushion to work with.

Yup.

> > Yes and no. Even before there was serious competition in the server
> > markets, Intel had better ASPs and margins on the mobile segment than
> > in server. AMD has had very good desktop market share starting with
> > the K6 (in fact, AMD has less market share today than it does when the
> > K6 was out). The problem was getting into the more valuable markets.
> >
> > Now today, AMD has made inroads into the server market, but you really
> > want to ask yourself, how much of a thread is this to Intel? The moron
> > I responded to seems to think that Intel is doomed and will be gone. I
> > seriously doubt AMD can hold > 25-30% of the x86 server marketshare.
>
> I don't think the other guy said that Intel itself is doomed. However,
> it does seem like he's saying that Intel will have a tough time catching
> up technologically.

Perhaps, perhaps not. I think Merom will do wonders.

> As for server marketshare, that's exactly the marketshare that they are
> aiming for right now, approximately 30%.

Let's see if they make it. AMD has never been able to hold on to their
market share gains or goals. They were supposed to hit 30% of the
general market, but they certainly aren't there now.

>As for how much of it is really
> a threat to Intel? I would assume it's quite a threat to it. Intel had
> 100% x86 server marketshare, 70% is a huge step down. It was accounting
> for a large portion of their profits due to the high margins and
> monopoly status.

I've talked with quite a few financial analysts about this, and since
Centrino came out, margins were higher for laptops ==> Intel made more
off laptops (since they have more volume there than in servers I
believe).

Ok, so are you saying that AMD will hit 30%? I want to see someone
here make an actual prediction, with times and stuff. I don't want "we
will accept whatever AMD says". Obviously companies have been known to
be wrong about product acceptance and marketshare projections.

> And there is no refuge in the laptop market either. Retail sales of
> laptops were 30% AMD just before Christmas, again up from nearly 0%
> before.

How much the laptop market is retail? I suspect not that much...

Moreover, AMD has always had retail laptops. Can you provide any
evidence that they had 1-2% of the retail market before? I was
certainly under the impression that they were doing around 5-10% of the
laptop market.

>The servers and laptops were where Intel had counted on making
> its profits, because it was making noises about how old-fashioned the
> desktop PC market was, and how the laptop was the way of the future.

Well, it is to some degree true.

> > They caught Intel at a very bad point in time WRT product line ups.
> > Intel has rectified this flaw; Bensley will be a reasonable start, at
> > least putting them in the right neighborhood for performance and
> > price/performance. It certainly won't get them ahead, but when
> > Woodcrest comes out, things will be interesting.
>
> Yeah but AMD didn't just catch Intel at a bad time, AMD created the bad
> time for them.

Yes, just like Intel did with Northwood for the K7 and how they did for
the K5/6. AMD's competitive pressure played a role, but Intel also
screwed up with Prescott.

> Intel was just doing fine with all of its old-fashioned
> chips, selling them without trouble, because nobody figured that they
> were old-fashioned yet. AMD had gone quiet for a few years while it
> worked on these technologies. It's not just the chips themselves, but
> also the manufacturing technology was was upgraded at the same time.
> Intel never thought about taking a breather and working on their future
> directions during the quiet time.

That's the funniest *** I've ever heard. You think the place where
"Only the Paranoid Survive" was just sitting on it's laurels? LOL.

>And now it can't even think about
> taking a breather, it's caught up in a full-scale bombardment, they have
> to spend time just shoring up their defenses.

Hurray, irrelevant military metaphors.

> > So, up until recently AMD had a strong technical advantage over Intel.
> > I think history has shown that when AMD can present a significantly
> > stronger product (say ~30-100% better, not just 10-15%, by whatever
> > your metric for better is), they tend to do well. The issue is that
> > historically, when AMD and Intel's products are very close (under 30%
> > difference) Intel has done very well. A large part of this is due to
> > marketing, channels of distribution, etd.
>
> I think that's right. In the last generation the Intel and AMD
> technologies were very close to each other. The Athlon XP and the P3
> were almost identical in IPC, with the Athlon getting ahead due to
> higher frequency. But in the current generation, AMD does hold the big
> performance and technology lead of greater than 30%.

Is it really greater than 30%? Are you talking about the desktop
market, laptop market, servers or what?

> > If AMD wants to be able to take and hold marketshare they need to have
> > a plan for dealing with Intel when they have no technical advantages.
> > They also need to be able to deal with Intel when they have technical
> > disadvantages. Intel is currently a year ahead of AMD in process
> > technology. They will have an advantage for that year or so, and then
> > AMD will likely end up ahead when they finally get 65nm worked out.
>
> Not likely, Intel had the same time advantage over AMD at 90nm, but it
> never worked out for them. The only thing that the lower nanometers give
> anyone nowadays is a cost advantage for manufacturing, but no
> performance advantages. It was starting to get obvious from the 180nm
> node on down that performance was not automatically scaling like it used
> to in the past.

No, I think 130nm had fine scaling. The problems started at 90nm.

> > The question is how will AMD fair this next year? I remember when
> > Intel had Northwood out, and AMD was still using 180nm parts...it sure
> > wasn't pretty and Intel took back all their marketshare and then some.
> > Of course, the cycle then reversed itself with 90nm.
>
> Yeah, but 180nm was pretty much the end of it for performance
> improvements.

AMD had a 130nm shrink of the K7.

> Intel always brings smaller process technology out six
> months or more ahead of AMD. When Intel transitioned to 90nm, AMD was
> still at 130nm for at least six months, but the performance was still
> increasing at AMD. Those days are over for "miniaturization is
> proportional to performance increases".

I know that.

> > Ultimately, AMD wants to be able to break the cycle, but I'm not really
> > seeing how they can.
>
> It seems to me that it already has broken that cycle now. It's been
> helped by physics: Intel can't use miniaturization as a crutch to help
> it get away from AMD anymore.

Perhaps...we shall see.

DK

.