Dumbity Dumb Article

[STL]

I feel like venting.

http://www.vanshardware.com/news/2004/03/040315_Bits_and_Bytes/040315_Bits_and_Bytes.htm

> For years we have detailed the weaknesses of Intel's NetBurst architecture

Sure, the Willy sucked. Then the Northy came along and dominated everything for a long time.

> it should come as no surprise that by perversely extending the Northwood
> P4's pipeline by over 50%, the Santa Clara chipmaker has created perhaps
> the worst dud in the history of microprocessors.

Please (iAPX 432? Or even IA-64, not that I agree with that?).

People made fun of the Willy/Northy 20-stage pipeline, and look where that got us.

> With Prescott's much deeper pipeline, the new Pentium 4's performance at
> any given clock speed is very often lower than its predecessor's
> (which is really hard to do considering that Prescott has twice the L2 cache!)

This is the utter bullshit that made me want to vent.

If you look at benchmarks, clock-for-clock the Prescott performs very closely to the Northy, within like 10%, and in some situations outperforms Northy.

Sure, the Prescott has twice the L2 cache, but that's good for maybe 10% more performance clock for clock (look at Northy 2.0A versus Willy 2.0 - not much difference going from 256 KB to 512 KB).

In fact, Prescott's similarity to Northwood reveals exactly what's going on inside of the Prescott's guts. Here you have a processor with a 55% longer pipeline, which directly affects its clock-for-clock performance, as well as the nastier branch misprediction penalty. The Prescott's other architectural improvements manage to make up almost all of this penalty.

> comparable with AMD's much more efficient Athlon 64.

"Efficiency" is meaningless, only final performance.

> Unfortunately, Prescott's heat output is so prodigious

Processors get hotter over time. Deal with it.

> This leaves Intel with a CPU that is much more complicated than the chip it replaces

What? That doesn't EVEN make sense.

> doesn't ramp as well under the same conditions

The longer pipeline will allow Prescott's frequency to continue climbing upwards.

> Why didn't the chipmaker simply "shrink" Northwood to 90nm, expand its L2 to 1MB, add SSE3

... because you can only go so far with a 20-stage pipeline.

> use fewer transistors than Prescott

Not appreciably. Most transistors in microprocessors go towards the cache (I don't remember exact figures).

> probably ramp better than Prescott.

Pure and utter bullshit. The only thing that makes Prescott different from their fictional-Prescott is the 31-stage pipeline.

> The answer is clearly that Prescott was a marketing driven product piloted straight over a cliff.

Prescott was not and has not been hyped as a performance demon. Extreme Edition, yes.

> Intel effectively spots the competition an IPC lead with NetBurst

"IPC" means nothing.

> Prescott has served a cold, heaping dose of raw humility to a company that many insiders have viewed as arrogant.

Meaningless. As much as I hate to say it, Prescott is not an NV30.

> Multi-core Banias/Dothan derivatives, a direction we championed in past columns, appear to be the future for Intel
> desktop processors after Tejas, Prescott's successor (and another item we exclusively reported).

"Go us!"

 

[Grazehell]

welcome back STL, do you still have your 2.4 northwood(I think it was)?

 

[BillLeeLee]

Hey STL.

Yes, I'd have to argue that some of the stuff in that article is 3 year old bunk. Netburst was weak early on, but when the clock speeds got to the point where it could actually take advantage of the 20 stage pipeline, it showed some muscle.

Regarding the whole Athlon efficiency thing: isn't this just because the Athlons have shorter pipelines, so they have to do more operations per cycle?

And does anyone actually remember hype over the Prescott? If there was major hype, this thing would have been called the Pentium 5; instead, Intel's deciding to confuse everyone with their P4E moniker. This was more like an underground ho-hum launch, probably due to the starting lackluster performance we've seen.

And Northwood uses less transistors yes, but the 1 MB cache of the Prescott makes up 48 million of the Prescott's transistors.

Prescotts I don't see as a viable option for myself as yet, but it represents a pretty advanced jump in microprocessor engineering. First chip to use 90 nm process, first chip to use strained silicon, and it also uses a 7 layer process against Northwood's 6 layers. This stuff will make the Prescott shine in future revisions (hopefully).

 

[STL]

[Grazehell]
> welcome back STL, do you still have your 2.4 northwood(I think it
> was)?

Hey hey.

Yes, I am still running Northy, named for its Northy-2.2A. Very soon I will be building Reason around a Prescott-3.4.

[BillLeeLee]
> Regarding the whole Athlon efficiency thing: isn't this just
> because the Athlons have shorter pipelines, so they have to do more
> operations per cycle?

Athlons have shorter pipelines; as a result, they are more powerful clock-for-clock. They also have other architectural differences which emphasize clock-for-clock power (e.g. a large but comparatively slower L1 cache).

> This was more like an underground ho-hum launch, probably due to
> the starting lackluster performance we've seen.

Yup.

Actually, there wasn't much hype over the Northwood either - only over the Willamette!

> First chip to use 90 nm process, first chip to use strained
> silicon

Heh, that brings to mind this:

http://www.legitreviews.com/Reviews/prescott_cooling_1.shtml

The energy lost due to the gaps in strained silicon is released in the form of heat. Thus leading to increased temperatures and the need for better cooling!

The only response to that is:

 

[Josh_B]

Originally posted by BillLeeLee
Regarding the whole Athlon efficiency thing: isn't this just because the Athlons have shorter pipelines, so they have to do more operations per cycle?

I agree with most of what you've said, but I will make one comment:

AMD designed the shorter pipeline because they KNEW it would increase efficiency. It wasn't as if they woke up one morning and decided 10/15 were the magic numbers

 

[Josh_B]

Originally posted by STL
[Grazehell]Athlons have shorter pipelines; as a result, they are more powerful clock-for-clock. They also have other architectural differences which emphasize clock-for-clock power (e.g. a large but comparatively slower L1 cache).

Hey - most of their team were former Alpha engineers. Is it any wonder their CPU looks similar in nature?

 

[Snugglebear]

lol, people read van's?

Anyway, as much as you hate them, IPC and design efficiency do matter. Alone, IPC is worthless, but taken as a part of a whole with clockspeed, memory bus, etc., it makes a considerable difference. Lengthening pipelines to gain extra clock cycles can only do so much for your architecture. Intel has just about hit their wall and that won't get better until the chemists work their magic. Even after the chemists do deliver, there's still only so far you can push - at some point you get more by widening your architecture than making it deeper and faster. The power drains alone are getting ridiculous with the long/fast arrangement, and, frankly, most users simply aren't consuming those clock cycles.

 

[Bomber]

Originally posted by Snugglebear

...most users simply aren't consuming those clock cycles.

Which is precisely why Intel designed Hyperthreading. They knew at the clock speeds they are running, the efficiency is crap.

 

[2phastPRO]

On the note about Efficiency.

It really depends on what you're looking at. The majority of the public is looking at mhz/ghz. So on that basis AMDs chips would be the more efficient ones because they do the job just as good if not better with less mhz/ghz.

The whole "Who makes the best chips" scenario turns nasty when you bring it up with tech junkies who know something about these chips or if you drop it in a forum such as this one.

In order to answer that question though...
As with everything... "It depends..."

 

[Snugglebear]

They're just tools. PPCs are great for low-power/heat applications, new Intels and AMDs are great when you don't mind having a heatsink that weighs as much as some dumbells. I'm just sitting back and enjoying the fact that what many pronounced as dead, both PPC and wide architectures, are coming back in full force and doing just as well as Intel's long/fast philosophy. The only thing I feel bad about is SPARC fading away.

 

[STL]

[Snugglebear]
> Anyway, as much as you hate them, IPC and design efficiency do
> matter. Alone, IPC is worthless

IPC is exactly as worthless as frequency. Only IPC * frequency matters.

[2phastPRO]
> So on that basis AMDs chips would be the more efficient ones

"Efficiency" doesn't matter.

 

[SLee]

Originally posted by STL

> use fewer transistors than Prescott
Not appreciably. Most transistors in microprocessors go towards the cache (I don't remember exact figures).

Actually, in the case of Prescott, this isn't true. Prescott is 125M transistors, of which 55-60M transistors are for the 1MB L2 cache. This leaves 65-70M transistors dedicated to the core logic, a massive increase from the ~28M transistors used in the Williamette and Northwood cores. The increase in transistor count from the larger L1 cache, architectural improvements, increased pipeline and 64-bit additions don't come close to accounting for the extra transistors, leaving a mystery as to what exactly several tens of millions of transistors are there for. So while Prescott is not impressive at the moment, a fully enabled Prescott (or Tejas perhaps) may become a very formidable processor.

 

[xonik]

Originally posted by STL
> With Prescott's much deeper pipeline, the new Pentium 4's performance at
> any given clock speed is very often lower than its predecessor's
> (which is really hard to do considering that Prescott has twice the L2 cache!)

This is the utter bullshit that made me want to vent.

Read Van's statement again. They said very often lower, which is true, not very lower often, which is not true.

If you look at benchmarks, clock-for-clock the Prescott performs very closely to the Northy, within like 10%, and in some situations outperforms Northy.

Sure, the Prescott has twice the L2 cache, but that's good for maybe 10% more performance clock for clock (look at Northy 2.0A versus Willy 2.0 - not much difference going from 256 KB to 512 KB).

In fact, Prescott's similarity to Northwood reveals exactly what's going on inside of the Prescott's guts. Here you have a processor with a 55% longer pipeline, which directly affects its clock-for-clock performance, as well as the nastier branch misprediction penalty. The Prescott's other architectural improvements manage to make up almost all of this penalty.

You're absolutely right about this, but like you tend to say, it's meaningless if there is not much of an actual improvement in performance.

> comparable with AMD's much more efficient Athlon 64.

"Efficiency" is meaningless, only final performance.

"Efficiency" is such a misnomer in this context. In terms of performance-per-watt-dissipated, AMD's Athlon 64 is more efficient. In performance-per-MHz, I would say that again, the Athlon 64 is more efficient. For you these things are indeed meaningless, but they aren't meaningless PERIOD, once and for all, amen, etc. For people with concerns with thermal management and power usage, efficiency is most definitely NOT meaningless.

> Unfortunately, Prescott's heat output is so prodigious

Processors get hotter over time. Deal with it.

...but if you're going to pay the price of cooling hotter processors, shouldn't you get more performance for your money?

> This leaves Intel with a CPU that is much more complicated than the chip it replaces

What? That doesn't EVEN make sense.

They're talking about transistor counts, which in my opinion, is an increase in complexity.

> doesn't ramp as well under the same conditions

The longer pipeline will allow Prescott's frequency to continue climbing upwards.

Theoretically. In practice, no one-- not even Intel--knows that this will happen. This is uncharted territory, remember. We are dealing with real physical limitations here.

> Why didn't the chipmaker simply "shrink" Northwood to 90nm, expand its L2 to 1MB, add SSE3

... because you can only go so far with a 20-stage pipeline.

This is an oversimplification of the problem, I agree.

> use fewer transistors than Prescott

Not appreciably. Most transistors in microprocessors go towards the cache (I don't remember exact figures).

This is mostly true, but of course Prescott also gained some 64-bit instructions and registers, which cost transistors. Also, the branch prediction circuitry costs some silicon and transistors.

> probably ramp better than Prescott.

Pure and utter bullshit. The only thing that makes Prescott different from their fictional-Prescott is the 31-stage pipeline.

Agreed, this is unsubstantiated, but I don't think you could prove otherwise, either.

> The answer is clearly that Prescott was a marketing driven product piloted straight over a cliff.

Prescott was not and has not been hyped as a performance demon. Extreme Edition, yes.

Yes, the writer overstepped his bounds on that comment.

> Prescott has served a cold, heaping dose of raw humility to a company that many insiders have viewed as arrogant.

Meaningless. As much as I hate to say it, Prescott is not an NV30.

Right now it is. The performance that Prescott has yielded can only be described as unimpressive, in comparison to Northwood and Clawhammer. So was NV30. NV30 was also a very hot little chip and required substantial cooling, which will be revealed when the new spiral heatsinks are used in Intel's forthcoming retail kits. The last NV30 similarity was that both chips were pioneers of new process technologies, and didn't really seem to benefit from them (except Prescott's small die size and the resultant cost benefits).

 

[STL]

[SLee]
> Actually, in the case of Prescott, this isn't true. Prescott is
> 125M transistors, of which 55-60M transistors are for the 1MB L2
> cache. This leaves 65-70M transistors dedicated to the core logic,
> a massive increase from the ~28M transistors used in the
> Williamette and Northwood cores.

I stand corrected.

> The increase in transistor count from the larger L1 cache

But that should be lumped in with the L2 cache.

> So while Prescott is not impressive at the moment, a fully enabled
> Prescott (or Tejas perhaps) may become a very formidable processor.

Yeah, and they've pulled this trick before with HT.

[xonik]
> Read Van's statement again. They said very often lower, which is
> true, not very lower often, which is not true.

True but meaningless - lower by 5% or whatever doesn't matter.

> You're absolutely right about this, but like you tend to say, it's
> meaningless if there is not much of an actual improvement in
> performance.

There will be, because of the latent ability to scale in frequency.

> For people with concerns with thermal management and power usage,
> efficiency is most definitely NOT meaningless.

Do you care how much heat your processor puts out? I don't. I slap a Swiftech on the sucker and forget about it. Do you care how much power your processor uses? I don't. As long as it works, I'm happy.

> ...but if you're going to pay the price of cooling hotter
> processors, shouldn't you get more performance for your money?

Prescotts at the same frequency as Northwoods are less attractive, duh. The whole point is that the Prescott can ramp up in frequency while the Northwood has hit a wall.

> Theoretically. In practice, no one-- not even Intel--knows that
> this will happen. This is uncharted territory, remember. We are
> dealing with real physical limitations here.

Yeah, they always say that.

> Right now it is. The performance that Prescott has yielded can only
> be described as unimpressive

It's not meant for performance right now.

> So was NV30.

No. The NV30 was twice as fast or more than the GF4. For Nvidia-only people, it was a no-brainer. It did not regress in performance relative to the GF4.

 

[xonik]

Originally posted by STL
[xonik]
> Read Van's statement again. They said very often lower, which is
> true, not very lower often, which is not true.

True but meaningless - lower by 5% or whatever doesn't matter.

You originally said it was bullshit. I showed you otherwise. I agree, it's largely meaningless, but it doesn't make the statement "bullshit," now does it?

> You're absolutely right about this, but like you tend to say, it's
> meaningless if there is not much of an actual improvement in
> performance.

There will be, because of the latent ability to scale in frequency.

Correction, latent potential. I would love to eat my words six months from now, a year now, but I can't do that just yet. I guess we all will have to see if that potential is realized.

> For people with concerns with thermal management and power usage,
> efficiency is most definitely NOT meaningless.

Do you care how much heat your processor puts out? I don't. I slap a Swiftech on the sucker and forget about it. Do you care how much power your processor uses? I don't. As long as it works, I'm happy.

Ah, you had to ask that question. YES, I DO care about heat output. I shouldn't accept that my computer heats up my room by ten, fifteen degrees. That's not what I paid for. I shouldn't accept increased power draw but the same performance. Also, increased power and thermal needs translates directly into the bottom line, both in terms of beefier hardware and beefier energy bills. It's money I could use for other things.

My petty situation is nothing compared to the concerns of businesses. Higher thermal and power demands starts to add up with thousands of computers being involved. This article was addressing the entire market, not just you and me.

> ...but if you're going to pay the price of cooling hotter
> processors, shouldn't you get more performance for your money?

Prescotts at the same frequency as Northwoods are less attractive, duh. The whole point is that the Prescott can ramp up in frequency while the Northwood has hit a wall.

No, Prescott should ramp up in frequency instead of hitting a wall. Should I question Intel's frequency ramping capabilities? YES. The 3.4E is still delayed, despite a paper launch occurring nearly two months ago. Should I ignore this shortcoming and expect the 3.6E to arrived as planned, in sufficient quantity? What about the 3.8E and 4.0E? Should I ignore the mumblings about a Dothan desktop arrival in the near future?

> Theoretically. In practice, no one-- not even Intel--knows that
> this will happen. This is uncharted territory, remember. We are
> dealing with real physical limitations here.

Yeah, they always say that.

So you know that Intel will overcome this problem in a timely matter then?

> Right now it is. The performance that Prescott has yielded can only
> be described as unimpressive

It's not meant for performance right now.

Are you referring to its lower cost then? That's not helping if Dell hasn't used the chips yet...

> So was NV30.

No. The NV30 was twice as fast or more than the GF4. For Nvidia-only people, it was a no-brainer. It did not regress in performance relative to the GF4.

You're right, it didn't regress in performance relative to the GeForce 4, but it did regress in overall performance standing, yielding the top spot to the Radeon 9700 Pro. Now if that's not important to you, I have nothing more to say...

 

[STL]

[xonik]
> You originally said it was bullshit. I showed you otherwise. I
> agree, it's largely meaningless, but it doesn't make the
> statement "bullshit," now does it?

Fine, I concede your point.

> Correction, latent potential.

That's what latent means.

> Ah, you had to ask that question. YES, I DO care about heat output.
> I shouldn't accept that my computer heats up my room by ten,
> fifteen degrees.

Come to think of it, my room is hot....

> beefier energy bills

A proc only draws as much power as a single 100W lightbulb.

> No, Prescott should ramp up in frequency instead of hitting a wall.
> Should I question Intel's frequency ramping capabilities? YES. The
> 3.4E is still delayed

Agreed (this is the proc I want, god damn it).

> So you know that Intel will overcome this problem in a timely
> matter then?

They've always succeeded before. ^_^

> but it did regress in overall performance standing

I don't think "regress" means what you think it means.

 

[xonik]

Originally posted by STL
> Correction, latent potential.

That's what latent means.

Okay, that's true, but it doesn't bring any more assurance of scaling abilities in the future.

> beefier energy bills

A proc only draws as much power as a single 100W lightbulb.

As opposed to a 60 watt lightbulb? Like I said earlier, it sounds insignificant and stingy from an individual perspective. What if a business changed all of its 100 watt lightbulbs to 60 watt bulbs? Ignoring the decreased brightness of the lighting, do you not think that there would be a significant energy savings?

> No, Prescott should ramp up in frequency instead of hitting a wall.
> Should I question Intel's frequency ramping capabilities? YES. The
> 3.4E is still delayed

Agreed (this is the proc I want, god damn it).Can't argue against that, I guess.

> So you know that Intel will overcome this problem in a timely
> matter then?

They've always succeeded before. ^_^

They will iron out the problem, I don't doubt that. But when will this happen? A month from now? Six months? Twelve months? The timeliness of this solution is of distinct importance, with AMD's processors arguably nipping at Intel's heels, or running astride them.

> but it did regress in overall performance standing

I don't think "regress" means what you think it means.

I define regress as "falling back to a worse position." In this context, yes, the NV30 fell into a worse position--second place.

 

[2phastPRO]

I fail to see why a processors "Efficiency" doesnt matter? Just a bit lost. Please explain.

 

[SLee]

Originally posted by 2phastPRO
I fail to see why a processors "Efficiency" doesnt matter? Just a bit lost. Please explain.

Because "Efficiency" is often used in terms of performance per CPU cycle. However, CPU cycles are not an endangered resource, that will run out, so it doesn't matter really how much performance per cycle you have (other than as a way to better understand the design philosophy of the CPU), so long as you have good performance per second. If you want to look at efficiency, it should be performance per second, performance per watt, performance per transistor, performance per mm^2...etc and not performance per cycle.

 

[STL]

Exactly.

> performance per second

This is the performance that matters to most people.

> performance per watt

This matters when you care about energy usage (e.g. it doesn't matter to me, it matters some to xonik, and it matters a hell of a lot in mobiles).

> performance per transistor

This doesn't really matter to anyone...

> performance per mm^2

But this matters to chipmaking companies. (They care about die size, not about transistor counts, since die size determines how many dice you can get from a wafer - at least, I think. I'm no EE.)

> performance per cycle.

This matters to AMD fanboys. :-P

 

[Snugglebear]

Originally posted by STL
> performance per second
This is the performance that matters to most people.

Most consumers don't even care that much anymore, and the vast portion of this performance isn't even processor-dependent, but rather I/O related (mem & disk). Intel and everyone else knows you want things to happen in under 100ms or else the user notices, which is why stupid programmers like us make spinning crap to show them something is going on. Anyway, it doesn't matter worth shit if you've got 4GHz CPU and an old 7200rpm ATA drive, the processor will just sit there twiddling away on NOPs.

> performance per watt
This matters when you care about energy usage (e.g. it doesn't matter to me, it matters some to xonik, and it matters a hell of a lot in mobiles).

It definitely matters in high-density environments, data centers, blade racks, etc. The amount of energy being consumed and vented by the chips these days is making it far more difficult to throw together large clusters.

> performance per transistor
This doesn't really matter to anyone...

More transistors = more heat. More transistors also mean more human design time unless you want to use design software, at which point you sacrifice immense amounts of performance and transistors.

> performance per mm^2
But this matters to chipmaking companies. (They care about die size, not about transistor counts, since die size determines how many dice you can get from a wafer - at least, I think. I'm no EE.)

It does mean higher yields. More transistors eat up more space, too.

> performance per cycle.
This matters to AMD fanboys. :-P

As I said before, combine it with other metrics.

 

[Kelledin]

Hey, as much as I hold Van's Hardware in contempt, he does have a point about Prescott's heat. The thing's a firebomb!

I know it doesn't bother most of us to pay for a few extra watts of power, or slap a big beefy HSF or exotic cooling on a CPU...but we're getting reports of power supplies burning out from Prescott's power draw, and a demo right here on [H]ardOCP of motherboard caps getting boiling hot just because of the demands Prescott places on the power rails. And wrapping a cooling solution around that is a real pain in the neck.

Besides...crap, if I'm going to pay for more watts, I expect to get something out of that. Even the difference between a 60W and 100W bulb is a sticking point if I don't get any bonus for using the 100W part.

Not to mention which, having these flaming bricks out where clueless or semi-clueless DIYers can just buy them and pop them in without a second thought is going to cause a lot of system burnouts, even if the CPU technically survives. I can see a lot of people getting pissed at Intel for that, even if those people should have known better.

What's strangest to me is, Intel didn't release the current stepping of Prescott for any terribly sensible reason that I can see. Intel didn't release it for added performance, Intel didn't release it for cooler running (DUH), Intel just released it to get more production headroom (more dies per wafer). It kind of bugs me that the mighty Intel with its dozen-or-so fabs going full-blast and its undeniably top-notch process technology feels the need to be stingy WRT production headroom. Especially risking a PR disaster and a black mark on its rep for platform stability.

Anyway, that's my rant for today.

 

[Dan_D]

Prescott may not make much sense right now. Initially the Pentium Pro suffered from the same problems that the newer ones did. They suffered from architechtual changes that made it perform worse with 16 bit code and faster with 32. But in Windows 95 we saw a performance hit.

But that processor matured and was revised. It became Pentium II which per clock wasn't as fast due to cache design. But the motherboards matured, then came Pentium III Katmai, then Coppermine, and finally Tulatin.

It's the same thing. These changes are stepping stones to larger changes that will come later.

Just like the Willamette sucked and didn't make sense. When Northwood came out and was revised it turned into a really powerfull and very desireable processor.

The same thing will happen for Prescott. Its just a stepping stone to bigger and better things. And as I understand it Prescott is almost as fast on benchmarks. So your average person wouldn't notice the difference between a Pentium 4 3.0C and a Pentium 4 3.0E. But eventually there will be gains to have.

We aren't processor or semi-conductor designers so theres no telling what the reasons are for the changes. Sometimes you have to back track to go another direction. Intel is one of the most advanced and sophisticated semi-conductor desginers/manufactures on the planet.

 

[§kynet]

Intel is one of the most advanced and sophisticated semi-conductor desginers/manufactures on the planet.

I would agree with that, and Intel has always introduced a new processor with seemingly terrible disadvantages relative to their refined previous generation.

I believe what may be different now is Intel has run into a wall in respect to clock speed. The whole Intel mantra for years is higher clock=better chip. It is such an easy way to judge processor "speed" just look at the bigger number. But what happens if say 4 gigs is the upper practical limit of clock speed for the next 2 years on any modern process? Where does that leave Intel? Their roadmap claims they will see 4 Ghz by years end. Even if they get there, what about 2005? Intel has painted themselves into a corner over clock speed and in their obsession to drive clocks higher and higher they have come up with a chip so hot even the most modern coolers have trouble coping properly, let alone motherboards and the current Prescott requires. This has never happened to any previous generation of Intel processor. Intel's new processor may have had a lower IPC relative to their previous but it didn't burn up power supplies, motherboards, and require a whole new case design. We all thought the original Athlon 462 was a room heater, but tests have shown that Prescott and full tilt can draw 150 watts or more. Does this sound reasonable to anyone? That is about the same power draw as 2 top speed Opterons.

Intel has sure put a log of faith in their engineers to get them out a bind. I am pretty sure Intel realizes that clock rates increases are not sustainable forever, hence their move to model ratings numbers. I believe this is only the first step to changing their one-track policy of higher clock=better speeds=more sales.

Ultimately transistors can only get so small and designs can only get so complex before diminising returns take over. The move to higher IPC IS the trend of the future, even Intel sees the writing on the wall. Prescott will be Intel's last chip designed with insane clocks in mind, I am pretty much certain of this.

 

[Snugglebear]

Yeah, I can't see clock rates going much further, especially at past or current levels of change. Moving to a larger word size is going to make frequency gains even more difficult. But worst of all is the shift to multi-core technology - it's something that's not to hard to implement and provides a big speed boost, but now software has to actually use them. If you've written multithreaded software, you know how painful this is. That most problems simply don't present a good parallel solution won't help much. Servers will always benefit from this due to the sheer number of processes and threads active on them, but desktop PCs with their single users...

 

[*foo]

OT

I have absolutely nothing to add to this thread. I just wanted to say Hi to Snugglebear and STL, as they're the only names I recognize around here from "the old days."

So Hi.

 

[Snugglebear]

I thought you said you lost your password over a year ago?

 

[STL]

It's *foo!!!

 

[rolo]

Heat output may not matter in certain environments. For example, if you are the type to buy some tall tower with lots of loud cooling devices installed, you won't care if your CPU is pumping out 40w or 140w. You'll have to shout in order to have a discussion near your computer, but it'll work fine no matter what.

Personally I prefer a quieter computer, one that doesn't announce itself to everyone entering the room. That's my preference though, and it doesn't really mean that the aforementioned preference ("more fans = more louder = more better") is wrong. The temperatures of my CPUs reflect this philosophy of mine, by the way.

However, I feel sorry for the engineers who are trying to design 2 and 4 processor servers and workstations around the upcoming Prescott-derived Xeon chips (Nocona, et. al.). How can you put 4 of these things in to such a small space and still be able to promise awesome "four nines reliability" (i.e. 99.99% uptime)? Especially if you want to cluster several of them near each other. Noise doesn't really matter for a server stuffed away in a rackmount somewhere, but how can you put 2 of them into a standard workstation tower and have it still be quiet enough that the user can still think? Many enthusiasts don't care about noise output but the majority of actual users do! At my job I certainly care that I can hear things other than the "whoosh" of my workstation's computer fans.

And what about Media Center computers? Ones that use higher-clocked Prescotts will double as space heaters, more so than current computers so.

I agree, however, that the article linked to in the first post is a terrible one, for numerous reasons.

 

[Snugglebear]

Again, microprocessors are tools designed to do certain things and thus better suited for various tasks than others. The problem with Prescott is simply that Intel borked it, miscalculated, fell victim to bad process problems (whatever, this isn't really relevant to the point), and as such the chip isn't very well suited for mass-market/OEM desktops, which is what Intel's major market is. Fortunately they've been rounding out their product line the past few years and have other standins that can fill in where Prescott doesn't work.

 

[Josh_B]

Originally posted by Snugglebear
Again, microprocessors are tools designed to do certain things and thus better suited for various tasks than others. The problem with Prescott is simply that Intel borked it, miscalculated, fell victim to bad process problems (whatever, this isn't really relevant to the point), and as such the chip isn't very well suited for mass-market/OEM desktops, which is what Intel's major market is. Fortunately they've been rounding out their product line the past few years and have other standins that can fill in where Prescott doesn't work.

...Only the paranoid survive

 

[IdiotInCharge]

Man, this is a nice argument. maybe this should be stickied to show just HOW you can get somewhere in a forum. Now, regarding the article, I enjoyed the excerpts, and actually from a more basic (non-hairsplitting) perspective found nothing outstandingly wrong with what was said. In fact, the theoretical processor described was exactly what I was hoping Prescott would be:

-More L2 cache, for the purpose of making the 20 stage pipline more efficient (therefore faster at all speeds) and for making hyperthreading far more efficient (which I think is very important)

-SSE3, giving the OS and applications instructions to help control hyperthreading such that threads can be complemented to be best executed in parallel

-A smaller process, which would yeild better clocks on the 20-stage design, while lowering temperatures

Number 3 did not turn out as well as hoped. Number one did not do what it could have for the 20 stage pipeline- it seems only to make up for the new stages decreasing efficiency. And the one thing I did not hear about until close to release, the longer pipeline, well, basically shot my hopes for prescott the moment I read it. I knew what that meant- and that is what we have.

NOW: few people really recognize here that Prescott is an immediate step back, compared to Northwood, and that is bad not because it's slower now, but because it amplified the P4's architectual disadvantages when compared to the K8's. Now, instead of looking foreward to a Prescott, I look foreward to a 939 pin athlon 64, because prescott has just too much of a step back from Northy in more ways than one, putting it far behind the A64's in the long run.

 

[STL]

> putting it far behind the A64's in the long run

You don't get it, as usual.

 

[Snugglebear]

I wonder what he'll think in the next year or two when A64s end up with more pipeline stages and decreasing IPC counts?

 

[xonik]

Originally posted by STL
> putting it far behind the A64's in the long run

You don't get it, as usual.

How do you know otherwise? Did you read the article at X-Bit Labs, showing the Northwood and Prescott at 4.2 GHz? The Prescott gained very little or no ground in terms of Northwood, even at 4.2 GHz, which should be the golden age of Prescott, right?

If current Northwoods are struggling to top the Athlon 64 right now, how is the slower Prescott going to do better? Don't say "by increasing the clock frequencies to __ GHz," because that statement is very flimsy right now.

 

[IdiotInCharge]

Originally posted by xonik
How do you know otherwise? Did you read the article at X-Bit Labs, showing the Northwood and Prescott at 4.2 GHz? The Prescott gained very little or no ground in terms of Northwood, even at 4.2 GHz, which should be the golden age of Prescott, right?

If current Northwoods are struggling to top the Athlon 64 right now, how is the slower Prescott going to do better? Don't say "by increasing the clock frequencies to __ GHz," because that statement is very flimsy right now.

Yeah, what he said. Really. Too bad I was too rushed to say this coherently in my post above, and am now too tired to care. lol. Anyway, the point is, I think, that Prescott is an obvious step backwards in several important ways, and therefore there is obvious cause for concern.

So, for those participating, what will Prescott have to improve upon, say in further steppings and compiler optimizations, in order to be foreward lookingly competetive in your eyes?

 

[§kynet]

Well, we will have to see what happens when Intel moves to Socket 775 (whatever that form factor is called.) Will 775 offer other improvements and changes versus Prescott? Don't forget AMD is not sitting still either. At current IPC considerations, an AthlonFX @ 3.0 Ghz is going to make life very miserable for Prescott even at 4.0+ gigz. That is what is so wierd about Prescott, it seems to have a zero future.

Do we all assume that Prescott is a proving ground for Intel before they move to 775 with a refined process?

EDIT: If you get a chance check out the Prescott article in the lastest issue of Maximum PC. They seem to think a lot of Prescott. Some of the assumtions they made really irked me:


"Prescott's architectural advances represent a significant, forward-thinking leap for processing technology. However, in terms of performance, you likely won't see any substantial performance benefits anytime soon. It will take between six and 18 months for applications to take advantage of the CPU's much longer pipeline and new instructions. But with that said, because of its price and scalability, this will likely be the CPU we're using by next year."

"At 3.4 Ghz, Prescott is simply idling; at this speed it's not really using the extended pipeline. Hence the lackluster performance we're seeing. As higher clock speeds begin to take advantage of the 31-stage pipeline, we'll see massive speed boosts."

Whoa wait a second here. Massive speed boosts? That is a hell of an assumtion to make! They know this how exactly? Because it has to be so? Wow. The whole article is basically a kiss-Intel ass glowing report and I was very turned off by it.

And what are they talking about when they say, "applications to take advantage of the CPU's much longer pipeline." How in hell can an application be coded or optimized for a processor's pipeline?

 

[xonik]

Originally posted by IdiotInCharge
So, for those participating, what will Prescott have to improve upon, say in further steppings and compiler optimizations, in order to be foreward lookingly competetive in your eyes?

An integrated memory controller would help somewhat. Possibly, silicon-on-insulator during fabrication. The rest of the problems are fundamental with the NetBurst architecture and will hopefully be addressed with Nehalem/Merom.

 

[STL]

[§kynet]
> check out the Prescott article in the lastest issue of Maximum PC.

Ugh, that article is reverse ass.

> forward-thinking

GAR. I hate that word. It means exactly nothing, like "arguably".

> take advantage of the CPU's much longer pipeline

That's like taking advantage of short, stumpy legs.

It doesn't make sense!

The longer pipeline is absolutely not attractive for any reasons but frequency scaling.

> As higher clock speeds begin to take advantage of the 31-stage
> pipeline

Meaningless!

> we'll see massive speed boosts.

Prescott will always be clock-for-clock slower than Northwood until things like FSBs increase, and maybe not even then. Duuh.

 

[xonik]

Originally posted by STL
"arguably"

"Arguably" means that you're willing to accept that many things are not absolute, and that other opinions exist besides yours.