Great news around the future of P4s and Hammers

[eden]

I like Anandtech's approach, they got much more information out of Comdex about the P4's future and AMD's.

For starters: Prescott DOES debut at 3.2GHZ. It also confirms my theory that Intel will have trouble with 0.13m heat-wise now that the 3.06GHZ outputs this much. A new stepping will probably heal some of it. Sadly the 0.09m Q4 03 rumor seems true, so AMD and Intel are both getting problems, and they may end up quite near of each others when releasing the 0.09 processors, by a quarter or two.
AMD officially confirms (though not publicly, soon) Barton will have a good 400MHZ FSB!
And as I had so many times linked you guys to, a website tested DDR400+400MHZ FSB and personally, the boost is quite amazing. They claim 0-15% performance boosts, compared to the weak 5% or 10% maximum off 333MHZ bus and mem.
VIA KT400A is also confirmed to be Dual Channel.
What still remains mysterious is Gigabyte's Hammer 800MHZ FSB claim plus Hyper Threading technology support.
nForce 2 is also for Hammer, this sounds weird as CH is single channel only, so Dual Channeling it would rather destroy the Opteron's greater advantage no?
Intel's Springdale chipset hoarde is revealed as well, with the 800MHZ FSB support. I still have a hard time getting how DDR400 will be accepted later on, unless DDR 2 is out by Q2 03.
What I do take interest in, is Canterwood. Seems there is no Rambus there, because it is labeled as the replacement for i850E, Intel's TOP of the line chipset. I find this weird, but I do like the fact they will optimize Canterwood to be stronger than the same feature Springdale one. I guess a better mem controller will make the top of the line chipset.
Finally, I found that funny, Anand tested the Hammer at 1.4GHZ and touching the heatsink while running 3d Mark 2001 (I think), it was "lukewarm", hehhe. It was also said it ran as good as a 2.2GHZ P4 or XP2200.

All this here: <A HREF="http://www.anandtech.com/showdoc.html?i=1752" target="_new">http://www.anandtech.com/showdoc.html?i=1752</A>

--
*You can do anything you set your mind to man. -Eminem

 

[Clockwurk]

Also from Anand...

The performance of the Athlon XP 64 at 1.4GHz was around the speed of a 2.2GHz Pentium 4 or Athlon XP 2200+

I can't help but think that this is a bad sign for AMD.. Initial (and AMD fanboy) predictions had a 1.6ghz Hammer around a 3200+.. Scaling from anand, that 1.6ghz will likely be a 2500+.. Hammer needs to scale very well, if AMD is to even stay in the game (a 2.0ghz Hammer will be approx. a 3100+)

 

[ritesh_laud]

For starters: Prescott DOES debut at 3.2GHZ.

Hmm, I think Anand said *at least* 3.2 GHz. I would expect that it would debut at 3.5 or higher.

It was also said it ran as good as a 2.2GHZ P4 or XP2200.

Let's see, an XP 2200+ is 1.8 GHz. So a 1.4 GHz CH performs like a 1.8 GHz Athlon. That's about a 20-25% gain in IPC (for 3D Mark at least). Pretty much what that AMD guy in Germany said back in March.

Ritesh

 

[eden]

If they said at least, wouldn't it fair to assume it starts there?
An approximate "least" clockspeed given would tend to indicate a need to start there first. Either that or the lowest end Prescotts start there, which would make sense.

Intel's strategy to have 800MHZ FSB and HT on 2.4GHZ P4s and above is what I very much appreciate, it's very caring, and a good strategy.

As for CH, it's pretty much a 28% increase in IPC, but we're also talking about 3d Mark. Look at the 3dMark scaling of CPUs, is it THAT much affected?
I would say if it gained this much under 3d Mark, it would be probably even more in other applications. Though nothing is certain so far.
I am starting to think AMD is considering some SMT technology as well, they did want to add more core improvements to compete.
Also, the 2.2GHZ P4 is 800MHZ higher than the 1.4GHZ Hammer. As it is, a 2.5GHZ Hammer is needed to compete the NW Bs with HT, i.e. 3.06GHZ.
So the future is still uncertain about the Hammer yet.

--
*You can do anything you set your mind to man. -Eminem

 

[SammyBoy]

Right... so let's just roll my thread into this one. What I see is that the Hammer will debut around 2GHz (or at least that's Anandtech's claim), I'd say it would have the speed equivalent to a 2600+ XP, or a 2.6GHz P4. Not too bad... not great, but if it scales well, then AMD might be able to rip off a huge speed boost like Intel managed over the last 8-10 months. If that were the case, come Prescott time, you might see the reverse of what we do today. Ahh... the joys of oneup-manship.

As to the P4... I was beginning to wonder if the P4 could keep scaling at it's breakneck speed. Yes, there are people out there with OC'd chips at 3.6GHz+, but with the current 3.06GHz putting out an amazing 82W of heat when HT is on (and what sane person wouldn't enable it?), that means that conventional cooling methods won't be enough. From a retail/oem standpoint, you can't have a CPU that puts out 85W+ of heat right now. The costs of the heatsink or other cooling measures (can we say watercooling from the OEMs?) would make things too much. Intel C1 stepping is reaching its thermal limit... and unless they swap that current IHS with a silver one, the CPU is going to need to be refined before it can get too much higher. My prediction: We will see a 3.4GHz NW with a 800MHz FSB out by July/August, and after that, a cooling/retooling period for Intel, followed by pre-Christmas flurry of CPU releases in October and November.

Somehow, I get the impression that AMD, while really in the doldroms, had a good idea that Intel couldn't keep raising the clock speeds at the pace it was for too long. I s'pose that we should've as well when we looked at the thermal characteristics of the latest P4 chips. And we should have assumed that HT would have increased thermal output, purely by the fact that more of the CPU is being utilized at any given time. Right now, P4s are running hotter than AXPs, but at a much higher speed. Now Intel is running out of room with the current P4 incarnation, and AMD really should be pouncing. Especially since it seems that the infrastructure for the release of the *shudder* "Athlon 64" is in place. Most board makers seem to be ready to flood the market at a moments notice... they just need CPUs to put in those boards. But alas, the AMD frontpage says availibility will be 1H'03. Ah well... not like I'm looking to upgrade anytime soon.

One last thing: In regard to dual-channel on the Claw. Last I heard, the northbridges of most the chipsets will still retain MCH capability, with the ability to turn off the on-die controller. That would explain how the NForce2 for the Hammer could still have dual-channel. It just wouldn't be integrated like the Opteron controller, and subsequently, only apps that rely on pure bandwidth without being affected by latency would benefit. Also, it could also explain how Nvidia got around the lack of a MCH on the chipset to share system memory with an integrated video card.

That's my 2 cents at least.

-SammyBoy

Some day, THG-willing, I shall obtain the coveted "Old Hand" title.

 

[eden]

I'd say it would have the speed equivalent to a 2600+ XP

That's nuts, it would mean a 2GHZ CH is as good as a 2.13GHZ AthlonXP! That's mad sad!

Anyway your hypothesis is agreeable, and I do see it that way as well.

--
*You can do anything you set your mind to man. -Eminem

 

[leonov]

We've already had indications that a 2GHz Opteron will give us SPECInt of 1202 and SPECfp of 1170. In 64-bit that could be about 20% higer still and that is only at 2GHz. Currently the XP 2400+ (at 2GHz) gives base figures of 808 and 693. Clawhammer is supposed to be about 15% slower than Opteron giving figures of 1021 and 994. These figures are then 26% and 43% faster respectively. Adding an extra 20% for 64-bit recompilation means we could end up with figures which are over 30% faster for SPECint (SPECfp doesn't normally improve with compiler technology as much as SPECint).

Also remember that an 800MHz Clawhammer beat a 1.6GHz P4 in Q3 and the scaling will improve at higher clock speeds. From this we may infer that at 1.6GHz it could match a 3.2GHz P4.

These figures don't prove anything except that applying linear scaling to inherently non-linear systems are unlikely to produce a valid result.

FWIW I believe that a 2GHz Clawhammer will be able to beat out a 3GHz P4 in most benches.

L

 

[bikeman]

... and the scaling will improve at higher clock speeds ...

NOT!!

Greetz,
Bikeman

<i>Then again, that's just my opinion</i>

 

[jclw]

Apparently the Clawhammer 1.4 at Condex runs as fast as a P4 2.2. Extrapolating, that means a Clawhammer 2.0 will run roughly as fast as P4 3.06. But AMD already has a 2800+ (at least on paper) that keeps up with a P4 3.06. AMD doesn't need another. It needs something that will keep up with P5/Prescott.

But then again all these numbers are speculation.

- JW

*PIII-800 @900 i440BX SMP and Tualeron 1.2 @1.74 i815*

 

[Clockwurk]

Intel C1 stepping is reaching its thermal limit



Agreed... I think Intel will sit on 3.06 for a while (AMD doesn't have a competitor). Then do two things; a process shrink and voltage drop. Intel might also make the IHS thinner (allowing faster heat transfer to the HS) or go to something more exotic (prob. not silver, perhaps all copper or even Carbon). Intel is fairly proficient in thermal engineering and shouldn't have too many problems overcoming the excess wattage.

 

[eden]

Technically you are wrong to extrapolate like that!
2.2-1.4=0.8
Thus, a 2+0.8=2.8GHZ performance.
AMD needs a 2.2GHZ minimum to keep up with the 3.06GHZ. However, considering the 2.25GHZ XP2800 keeps up with the 2.8GHZ, a 2.45GHZ would keep up with the 3.06GHZ. If a 2.2GHZ CH can compete the 3.06GHZ, a 250MHZ boost equals the IPC boost the CH has over the AXP, and thus the scaling in fact is very weak here.
Either the AXP has weird scaling, or it's the effect that the nForce 2 plus higher bus boost that has caused the CH's IPC advantage to be rather diminished. Therefore, I would say a DC DDR nForce 2 Hammer combination, would in fact make a 2.2GHZ perform as good as a 3.2GHZ HT enabled P4, and would then, I guess, make your extrapolation be right.

--
*You can do anything you set your mind to man. -Eminem

 

[imgod2u]

Performance disparity doesn't scale linearly. I.e. the difference in clockrate for similar performance doesn't remain the same. An Athlon at 1.4 GHz may perform similar to a P4 at 1.8 GHz, but an Athlon at 2.25 GHz performs similar to a P4 at 2.8ish GHz. If you want to talk just 3dMark, the correct math would be 1.4/2.2 * 3.06 suggesting that a 1.94 GHz Hammer would be similar to a 3.06 P4 (keep in mind the 2.2 has a 100MHz FSB while the 3.06 has a 133MHz FSB and the 3.06 has HT while the 2.2 does not).

"We are Microsoft, resistance is futile." - Bill Gates, 2015.

 

[eden]

You are forgetting the AthlonXPs at 2600 and above have 333MHZ FSB support so nForce 2 DC DDR optimizations come in.
To me, an AXP vs a P4 100MHZ FSB, has to have a 400MHZ gap to tie. With the new FSB, it takes 300MHZ gaps, or even less. The new FSB and board gives that 400MHZ gap back.
Though I have a hard time actually seeing how the heck do these CPUs actually have a bigger gap and yet tie the 2.8GHZ which is 550MHZ higher... Dunno but either the board and FSB help THAT much or the AXPs had some weird scaling between 2GHZ and 2.25GHZ.

Again remember that the AXP FSB and chipset make a difference, a big one.
And the P4's FSB and chipset also do, therefore the two side's optimizations in the end negate the advantage one has over other, so the gap remains nearly the same in the end, except for the weird 2.25GHZ tie-in against the 2.8GHZ.
I was noticing occasionally linear scaling problems between the 2.8 and 3.06, but I could be wrong.

--
*You can do anything you set your mind to man. -Eminem

 

[Schmide]

Theoretically, the scaling of a processor should follow a logistic equation (a flat S). Up to the speed of the FSB you get improvements for the rate of memory transfers. From there the multiplier scales performance aggressively upward as the speed of raw calculations improves. At some point the memory system is unable to feed the processor and thusly the curve flattens out.

Dichromatic for your viewing plesure...

 

[leonov]

According to Fred Weber tech uber guru at AMD

"The controller scales memory latency, so as the CPU and HyperTransport link speeds improve, memory latency does as well,"

from http://www.eetimes.com/story/OEG20011016S0092

He should know.

Basically because the mmeory controller is on-die and is clocked at CPU speed the stages of the memory controller get faster with CPU speed. The latency of the memory access still takes the same amount of time so it can never decrease to 0 but rather asymptotically towards the memory latency.

L

 

[imgod2u]

The Athlon may have gotten an FSB boost at 2.13 GHz and above, but so did the P4. So in reality, the scaling of both processors is still on track. You'll notice that before the boost in FSB for processors above 2 GHz, the disparity was shrinking. Now, it's grown somewhat. As both processors scale, I suspect that the disparity will shrink yet again until Barton comes out at 2.4 or 2.5ish GHz and receives another FSB boost. The point is, the clockrate disparity does not stay the same.

"We are Microsoft, resistance is futile." - Bill Gates, 2015.

 

[eden]

You were repeating what I just said!
I said while the improvements were on AMD's side recently, they also occured for P4, and both had similar increases of performance, thus the gap is pretty much negated and back to normal. Now you are infering that the Palomino core at 2GHZ had started gaining more performance in the gaps, thus explaining why the 2.25GHZ has an easy time negating the 550MHZ gap. If that's true, then it helps understanding the odd concept of scaling and deltas of the scaling.
Of course as mentioned, the Barton will have a 400MHZ FSB, attaining finally the first Pentium 4 speeds, and since it starts with such FSB at 2GHZ+, it should no doubt have a great deal of perfomance boost. I already had posted a url to a site which has benchmarks about an AXP with 400MHZ FSB and DDR, plenty of performance extracted out of it, and at its time, it was a low clock speed AXP! There was almost a 50% increase in Quake 3, if I am not mistaken. I just hope the improvement will be significant with the nForce2 combinations. It just might allow some small competition until ClawHammer. It will definitly not be the performance crown champion, as SSE2 is a huge restriction. I'm just glad AMD is proving some that the K7 is definitly not a 'dead beat horse', and that you can still get the most out of any core out there, proving that the P3 Tualatin has so much potential that was omitted.

--
*You can do anything you set your mind to man. -Eminem

 

[imgod2u]

No, I'm saying that the increase in clockspeed disparity for similar performance is normal. And that the reason the disparity was decreasing before was because of the FSB boosts for the P4 while the Athlon stayed on the same FSB. So no, there isn't a set "Athlon will perform the same on average as a P4 of x MHz more" because that x will increase or perhaps decrease as the clockrate of both processors increase. So far, there seems to be an indication that the disparity will increase, however, you can never be sure. Eventually, the Athlon will be limited by its L2 cache and won't scale as well as the P4 with its superior L2 cache.

"We are Microsoft, resistance is futile." - Bill Gates, 2015.

 

[bikeman]

First of all, my excuses for my somewhat short and fierce answer I gave you. I was in a little hurry in that moment, but I certainly wanted to reply to that. Because ...
Secondly, I still cannot agree with what you say. My statement: The Hammer will scale less than linearly as every single processor does and ever will (though that last part could be overly courageous to state ...). Even if you consider a perfect memory controller, I am quite sure that my statement stands true. It is actually a quite basic reasoning that's behind. First Hammer: 333MHz Memory/800 MHz clockspeed. This hammer: 333/1400 = less. Memory becomes more of a bottleneck, 'cos in each clockcycle, less data can be transferred from RAM to the caches or the registers of the CPU. At 2000 MHz, you get 333/2000, so even less is at it's disposal. Even if the increased clock speed makes the memory controller more efficient (though that seems rather unlikely), the thing I just explained would overrule the gained efficiency. That's what I think of it. If you, or somebody else can claim the opposite ... Please go ahead ...

Greetz,
Bikeman

<i>Then again, that's just my opinion</i>

 

[eden]

Now even hardware news websites are claiming K8 will have an FSB, and one of 800MHZ. I have a hard time beleiving it, especially as Intel is supposedly the first to reach it by next year, and that the K8 does not have some QDR algorithm.

An FSB architecture explanation of K8 will be necessary.

--
*You can do anything you set your mind to man. -Eminem

 

[jclw]

The K8 Hypertransport bus runs at 800mhz which is probably what they are talking about.

I wonder what kind of memory the K8 at comdex had?

The latency point people brought up earlier is interesting. I don't think the memory controller will run faster on a chip with a higher multiplier (and therefor reducing latency). I think it'll be on the "non-multiplied" side of the chip, if that makes any sense.

*PIII-800 @900 i440BX SMP and Tualeron 1.2 @1.74 i815*

 

[bikeman]

What you imply is that the Hammer would have something like an on-die FSB, am I right? If the memory-controller would have a multiplied part (i.e. running at the same speed the rest of the chip is), a higher multiplier implies more time for the mem-controller to calculate stuff, like predict upcoming memory reads. I mean, there are more clockticks available to get a more intelligent memory-controller. But as I stated above, I estimate the chance that this will be implemented, quite low ...

Also a note on your signature ...
A Tually 1.2 @ 1.74 ??? How do you do that? I suppose you cool quite aggresively ... With the stock cooler, a friend of mine doesn't get over 1500 MHz ...

Greetz,
Bikeman

<i>Then again, that's just my opinion</i>

 

[vacs]

Now even hardware news websites are claiming K8 will have an FSB

I don't want to sound rude but after the release of the 486 DX 66 I never again heard of a PC CPU without a FSB... How do you want to synchronize your memory interface if your bus is running @ 2GHz real speed?

If I remember correctly the Hammer running @800MHz previewed earlier this year at tecchannel.de had a FSB of 200MHz (therefore 400 effective MHz).

 

[jclw]

I was just browsing through the AMD whitepapers and it looks to me like both the hypertransport link and the memory link will both run at 800mhz (both part of the "crossbar" link system), creating, essentially, an "on chip FSB".

I'm sure my little Tuallie will go faster but it's being held back by my memory - right now I just have a 64mb stick of PC100 but I've got a 256mb stick of PC150 CL2 on the way . I'm going for 1800 (150FSB) on air. For cooling I'm just using an old AMD cooler that I peeled off a Tbird-900 two years ago or so, with the same old thermal goop. The Asus monitoring utility says it idles at 43C and runs at 48C loaded. It's one of the new tB1 stepping cores, made the 29th week of this year in Malaysia.

Also, I did a VID pin mod so it cold boots at 1.65v instead of 1.50v

*PIII-800 @900 i440BX SMP and Tualeron 1.2 @1.74 i815*

 

[eden]

No rudeness taken, I know what I said.
It's not about not having an FSB, it's more like "how do you explain" the FSB functioning of the K8?
How would an FSB on K8 actually run at 800MHZ?
Why do they call HT the FSB, if that's just like a PCI to South or Northbridge connection, just faster?
This whole rumor about no FSB, or an FSB at 800MHZ just plain mixes me.

--
*You can do anything you set your mind to man. -Eminem