Power5 with hyperthreading

[Catfish_Man]

Originally posted by iwantanewmac
I bet this all isn't going to take place this year.....
First we get another lame G4 speedbump with some other minor updates that almost no1 uses.

Seems fairly likely. A bump in early Q3 using 1.6GHz 7457s on the not officially recognized 200MHz bus (the way the 7455's Apple used had 167 before Motorola officially admitted it could), DDR400, maybe USB2, and maybe a new graphics card and/or AGP8x. Maybe a dual 1.33GHz 17" Powerbook (note: I rather doubt that, but it might be possible with the '57). Then in January they announce 1.4/1.6/1.8GHz PPC 970 based PowerMacs shipping in February.

<disclaimer> Everything above is, at best, an educated guess. </disclaimer>

 

[ffakr]

Originally posted by Catfish_Man
Seems fairly likely. A bump in early Q3 using 1.6GHz 7457s on the not officially recognized 200MHz bus (the way the 7455's Apple used had 167 before Motorola officially admitted it could), DDR400, maybe USB2, and maybe a new graphics card and/or AGP8x. Maybe a dual 1.33GHz 17" Powerbook (note: I rather doubt that, but it might be possible with the '57). Then in January they announce 1.4/1.6/1.8GHz PPC 970 based PowerMacs shipping in February.

I think Apple is acutely aware of the G4 problems and they'll get a 970 based machine out asap. I've no doubt that the machine is being developed with whatever pre-production processors that IBM can spare.
Maybe I'm just being an optimist.. but I expect a brand new 970 based Powermac sometime this year.
They can't afford to release a next gen machine one day later than necessary.

 

[strider42]

Originally posted by ffakr
I personally wouldn't be surprised if a .13 micron or .09micron Power5 with a single core becomes the Mac chip in late 2004 or early 2005... but who knows. This is also pretty much impossible to say since the Power4 isn't put in anything but IBM servers. There is no $/1000 unit price, no raw or even retail price. Any figure generated for a Power4 module is the price for an IBM upgrade part for a high end server. I'd wager that if you wanted an IBM 'certified' Xeon, it'd cost a fortune too. I remember pricing additional Pentium 90 cards for old IBM 720 servers and they were around $1000... back when a P90 was a $30 part. Unfortunately, we needed to buy the card with with the CPU soldered onto it. I think the real test of whether the Power5 makes it to the desktop is whether or not it includes Altivec. Apple would demand such support. IBM didn't think it was too important back in the day, but it has proved to be powerful and it would be useful in server situations.. so we'll see.

It is absolutely impossible that the power4 or power5 will be in a desktop machine. absolutely, utterly impossible. Its not aimed at the desktop market. Its bulit for absolute reliability, and is thusly very expensive and gives up performance because of it. The power4 and power5 are absolutely NOT desktop chips, and no one in their right mind would use them as such. the 970 IS a desktop chip, designed to be so. if a power5 derivitive is made, that will be a desktop chip, but these are NOT the power4 and power5.

 

[ddtlm]

strider42:

It is absolutely impossible that the power4 or power5 will be in a desktop machine.

I know what you're trying to say, but technically it is not impossible... just exceedingly unlikely. They could make a $10k Power4 desktop which is evenly divided between heatsink, powersupply, and "other stuff".

gives up performance because of it

Yes, the whole design does sacrifice some performance, but there is no evidence that it will be any different for the PPC-970. As far as I can tell the only reasons why a PPC-970 will sometimes outperform a Power4 are: the 970 will be clocked higher, and it will have AltiVec.

 

[ozubahn]

Originally posted by ddtlm

Yes, the whole design does sacrifice some performance, but there is no evidence that it will be any different for the PPC-970. As far as I can tell the only reasons why a PPC-970 will sometimes outperform a Power4 are: the 970 will be clocked higher, and it will have AltiVec.

There is evidence, and you just pointed it out. The higher clock rate implies either a longer pipeline or thinner gate oxide (or both), the latter made possible by the reduced reliability requirements. These changes, along with a single-core die and vastly cheaper chip packaging, are what make the 970 and the POWER4 entirely different beasts. We can't guarantee that some hobbyist won't pick up an old POWER4 MCM at a swap meet years from now and stick it in a desktop box, but other than that, it isn't going to happen. The POWER5 may be a better hope, since IBM has suggested that it will move downmarket, but it still isn't likely.

 

[ddtlm]

ozubahn:

There is evidence, and you just pointed it out.

The higher clockrate is more likely possible for the same reasons for Xeons and Athlon MP's always are sold slower than their stand-issue brethren: lower standards, less testing. The effect of the smaller, single-core die could also be significant because of much better odds of the whole thing coming through with reduced speed-limiting defects. Power4+'s currently clock at up to 1.45ghz, not too far off the PPC-970's 1.8ghz, which obviously isn't even here yet.

but it still isn't likely

I never claimed otherwise.

 

[hvfsl]

I think we should dump moto and forget about IBM and go to Intel. The P4 is basically a risc chip that has a unit that converts x86 instructions to RISC and back again. What we need is a modified P4 that is PPC compatible, since Intel is the biggest chip maker in the world, this should not be too hard for them.

We would then always be able to keep up with the Windows world. The problems people have with Intel should also be left behind because Intel doesn't like M$, it is trying to find OS's that can replace Windows. It looked at getting BeOS at one point.

 

[firestarter]

Originally posted by ffakr
true... except I belive IBM said they have no plans for a multi-core 970.IBM isn't claiming 4x performance increase from HT... they claim that their implementation is MUCH better than Intels and that a single core HT Power5 will really act like 2 processors. They claim 80-100% increase with their HT... proof remains to be seen. IBM has also said the core has been reworked so there are also performance improvements totally unrelated to HT.
Intel's HT, OTOH, seems to produce anything from a 20% increase to a 10% decrease in performance.

The trouble with all these Intel comparisons is that Intel is a (quicly) moving target. The 970 may be competetive against the current range of Pentium 4s, but Intel will be another step ahead by the time the 970 is released.

From Anandtech: http://www.anandtech.com/showdoc.html?i=1790

"Intel's Prescott core will be their first CPU built on their new 90nm process, and is due out before the end of this year.

The updated Intel NetBurst micro-architecture comes in the form of an 800MHz FSB and a larger L1 data cache of 16KB, up from the 8KB of the Willamette and Northwood cores.

The improved Hyper-Threading technology in Prescott comes in the form of two new Hyper-Threading specific instructions, that are a part of the 13 total new instructions that made it into Prescott.

We've been mentioning for a while that Prescott will feature a large 1MB L2 cache, and Intel has publicly confirmed this at IDF. The 1MB L2 cache will bring Prescott up to around 100 million transistors, rivaling some of the most complex GPUs in transistor count.

The 90nm process combined with additional enhancements to Intel's manufacturing processes will allow the Prescott core to scale to the 4 - 5GHz range before it will be replaced by Tejas, which we mentioned earlier today. "

There is NO WAY that a 1.5 to 2GHz 970 is going to be faster than a 4 or 5 GHz Prescott (which also incidentally adds further Altivec style instructions, including specific support for video encoding).

 

[DharvaBinky]

Originally posted by spinner
Only the 3.06GHz P4 and later have Hyperthreading technology. Code must be specifically written to take advantage of it, and so far the only software I know that takes advantage of it is Windows XP. There could be others but I am not aware of any.

This is, in fact, incorrect. Software does not have to be rewritten to take advantage of SMT (Simultaneous MultiThreading the "real" name of HyperThreading). All software which runs on more than a single thread automatically takes advantage of HyperThreading. An example (and an anti-microsoft jab *grin*):

My team has written a large web app for universities to manage student flow that runs on Microsoft technologies such as SQL Server 2000. We're running our systems on Dell PowerEdge 2650 servers which feature Dual P4Xeons (with HyperThreading). Although the machine has *2* physical processors, Microsoft requires that we buy *4* processor licenses for SQLServer since the system appears as a quad processor machine to the Operating System. <OUCH> that put a big hurting on the 'ol budget there... SQLServer ain't cheap.

Apple and others need to watch out for this if/when IBM introduces the PPC980 (power 5 decendant), since many server software packages are licensed per processor (unlimited users). Windows 2003 purportedly solves this problem by recognizing HT processors, however... So, hopefully, Apple will learn from this MSNAFU and do it right in OS X the first time.



Dharvabinky

 

[Frobozz]

I wish you were joking.

Originally posted by hvfsl
I think we should dump moto and forget about IBM and go to Intel. The P4 is basically a risc chip that has a unit that converts x86 instructions to RISC and back again... We would then always be able to keep up with the Windows world...

UGH. lol.

 

[ryan]

Originally posted by wdlove
I was enthusiastic after reading the article. Now I'm depressed, is there any hope for the future like MWNY (that is if Apple show up)?

Apple will release a 970 based PowerMac when a 970 based PowerMac is ready. In case you've missed it Apple is now releasing products when they're ready and not waiting for some arbitrary date on the calendar, i.e. the Macworlds.

 

[ffakr]

Originally posted by firestarter
The trouble with all these Intel comparisons is that Intel is a (quicly) moving target. The 970 may be competetive against the current range of Pentium 4s, but Intel will be another step ahead by the time the 970 is released.

first off... I compared IBMs HT implementation to Intels HT... I didn't comprare the 970 to the P4. You are taking my comment out of context.

"Intel's Prescott core will be their first CPU built on their new 90nm process, and is due out before the end of this year.

The updated Intel NetBurst micro-architecture comes in the form of an 800MHz FSB and a larger L1 data cache of 16KB, up from the 8KB of the Willamette and Northwood cores.

wow, another P4 built on the same core with upgraded stats. 8K of additional L1 cache (btw, I'd expect this to be instruction not data cache.. but what do I know).

The improved Hyper-Threading technology in Prescott comes in the form of two new Hyper-Threading specific instructions, that are a part of the 13 total new instructions that made it into Prescott.

We've been mentioning for a while that Prescott will feature a large 1MB L2 cache, and Intel has publicly confirmed this at IDF. The 1MB L2 cache will bring Prescott up to around 100 million transistors, rivaling some of the most complex GPUs in transistor count.

So intel continues to build ever larger processors in the pursuit of ever diminishing returns. The Intel x86 project is quicly becoming the stereotypical 'American Musclecar'. Big, powerful, and not especially efficient. Even at .09 micron, I'm sure this will be a big chip... certainly big by the standards set by other .09 micron chips. I guess we'll have to wait to see if Intel runs into problems cooling such this sucker considering the number of gates in such a dense array.

The 90nm process combined with additional enhancements to Intel's manufacturing processes will allow the Prescott core to scale to the 4 - 5GHz range before it will be replaced by Tejas, which we mentioned earlier today. "

There is NO WAY that a 1.5 to 2GHz 970 is going to be faster than a 4 or 5 GHz Prescott (which also incidentally adds further Altivec style instructions, including specific support for video encoding).

[/b]
Here is where your argument falls apart. The Prescott core will scale to 4-5 GHz. That is the speed predicted over the life of the core. The 970 will likely DEBUT at 1.5-2GHz, and the Prescott will DEBUT lower than 4-5GHz, but it may eventually scale that high. This is Extremely important. We don't know but a 1.8 or 2GHz introduction of 970 on .13 micron will very likely scale north of 3GHz on .09 micron. Will a 4GHz Prescott be faster than a 3GHz 970? We have no idea. The current SPEC seems to indicate that the 970 would give it a damn fine run for its money.
Another thing to consider... do you expect a 5GHz P4 core to be 25% faster than a 4GHz P4 core? You don't expect these ultra high clock speeds to show substantially diminishing returns as the CPU runs away from the speed of the components?
IMHO... the key to this decade isn't 10GHz processors, it's more intelligent design. IBM is on the right track with what appears to be a much better SMT implementation and a smaller, more efficient cpu core. Even Sun seems to get it. They expect to have massive SMT and multi-cores per die... Sun hopes to have 8, 16, maybe 32 threads per processor die by 2005.

 

[strider42]

Originally posted by ddtlm
strider42:


I know what you're trying to say, but technically it is not impossible... just exceedingly unlikely. They could make a $10k Power4 desktop which is evenly divided between heatsink, powersupply, and "other stuff".


Yes, the whole design does sacrifice some performance, but there is no evidence that it will be any different for the PPC-970. As far as I can tell the only reasons why a PPC-970 will sometimes outperform a Power4 are: the 970 will be clocked higher, and it will have AltiVec.

Is it evidence enough that IBM THEMSELVES have talked about the differences between the 970 and the power4 and talk about how the 970 is actually faster in some areas because it isn't making compromises necessary for a server chip. Unless apple gets into the high end server business, it is utterly impossible this chip will ever be used by apple. OK, you are right its not technnically impossible, but the likelihood is just about 0. You can think what you want, but the 970 is too large, too expensive and makes too many compromises to be useful to apple. The 970 would be a better chip for apple in all possible areas except very high end servers. the 970 would be better as a workstation chip, a desktop chip and a mobile chip. There is nothing about the power4 that would even be remotely interesting to apple aside from the dual core aspect, which helps make it very expensive (among other things)

The 970 will be faster not because of altivec and higher clocks (althought hose certainly help), but because its not making compromises for reliability. IBM themselves have talked about this. If apple was going to use the power 4, they could have used it for nearly two years now. Its just not going to happen. its going to be the 970 and possibly a derivitive of the power5 if such a beast appears.

 

[ffakr]

Originally posted by strider42
It is absolutely impossible that the power4 or power5 will be in a desktop machine. absolutely, utterly impossible. Its not aimed at the desktop market. Its bulit for absolute reliability, and is thusly very expensive and gives up performance because of it. The power4 and power5 are absolutely NOT desktop chips, and no one in their right mind would use them as such. the 970 IS a desktop chip, designed to be so. if a power5 derivitive is made, that will be a desktop chip, but these are NOT the power4 and power5.

First off... I never said the Power4 would be in a Mac. Drop that part, it's a non-issue.

Your argument doesn't make any sense though. Just because it is reliable doesn't mean it won't work in a desktop. Just because it is currently prohibitivly expensive doesn't mean it won't work in a desktop (though market forces would make it impractical).

There is nothing in the architecture of a processor that makes it inherently expensive to create. Cost is related to R&D, Demand, Supply, Production costs (size, yield,...). It isn't related to the way in which the gates are arranged in a litho mask.
Server processors do tend to have thicker gates and interconnects to increase reliabilty, but even this is a minor extra cost. They are also sorted to a higher standard than desktop parts, increasing testing cost and decreasing yeild.
Once the 'maze' of the architecture is laid down, the trivial part would be shrinking the 'walls'. It would be a LOT easier than maintaining an architecture and a derivative of that architecture.

Now, look at my argument again. What makes more sense, making a Power5 for the desktop/small server or making a Power5 and yet another architecture based off of the Power5? The cost associated with making a desktop Power5 is MUCH lower than making a derivative architecture... and the dev time would be much shorter too.

As I said... I don't know if a Power5 will end up in a Mac and you certainly don't either. IBM is hinting (over a year out) that this processor will run in smaller hardware than the Power4 was designed to run in. The Power4+ (smaller die) is being put into smaller servers already. Add to this the fact that NO ONE knows what a Power4 really costs because the only info you have is the cost of an IBM aftermarket upgrade part (which is an IBM branded module with a big L3 cache)... not a raw chip price. IBM doesn't sell them to anyone else, so we have no idea. We have no idea how increased production and increased supply for a much larger market would affect the production cost of a chip like the Power4+ and we certainly don't know how it would affect a processor like the Power5... which just got powered on 2 weeks ago.

I think the key to whether we may see a real Power5 in a Mac would be whether or not IBM includes Altivec in the Power5, and what the transistor count turns out to be.

One more thing... if IBM does design the Power5 to be an option for the Mac, the increased production will drive down per unit cost. This means larger margins and more competitive prices for their Enterprise hardware.... their largest business unit.

 

[ffakr]

Originally posted by strider42
The 970 will be faster not because of altivec and higher clocks (althought hose certainly help), but because its not making compromises for reliability. IBM themselves have talked about this. If apple was going to use the power 4, they could have used it for nearly two years now. Its just not going to happen. its going to be the 970 and possibly a derivitive of the power5 if such a beast appears.

I believe that Power4+ IS as fast or faster than the 970 when you don't consider the MHz increase.

from SPEC.org
IBM pSeries 650 Model 6M2 (1450 MHz, 1 CPU) 1 909 935
IBM eServer 655 Model 651 (1300 MHz, 1 CPU) 1 822 848

estimated SPEC for the 970 at 1.8 GHz are about the same for Int and slightly higher for FP. IBM says the 970 numbers will likely go up as they near production, but I think a 1.45GHz Power 4 would hold its own against a similarly clocked 970.... don't you?

 

[scem0]

Originally posted by howard
all we can do is wait... lets just hope we don't have to for very long!!

yeah, we won't be seeing this chip in a mac for a looooooong
time, if ever... so I think it would be best if we all just tried not
to relive the whole 'g5 is coming!' thing again.

'The power5 is coming!'

 

[MacRETARD]

Originally posted by spinner
Only the 3.06GHz P4 and later have Hyperthreading technology. Code must be specifically written to take advantage of it, and so far the only software I know that takes advantage of it is Windows XP. There could be others but I am not aware of any.

The Intel XEON chips also have HT. And no you dont have to specifically write code to support it. If you are running more than one thread then it is automatically supported. To the OS it appears that the machine really does have 2 cpus. The performance advantage is not as good as 2 real cpus though. According to most reviews and benchmarks in some situations you will see a 25% performance increase. The big advantage is that the OS will run much smoother since tasks will have the advantage of the virtual dual cpu. This means launching apps and switching back and forth between apps will be much smoother.

 

[applejilted]

The danger with IBM is, IMHO, that IBM is also a a major server manufacturer and as such wants to incorporate its chips in its boxes. Apple has a modestly successful server product (Xserve with OS X server software) will IBM want to supply Apple with their Power5 chip (albeit not the multicore variety) so Apple could compete with IBM in what IBM considers a crucial market ? This may seem trivial now as Apple is a mere minnow in the server market comapred to Big Blue or SUN but don't think for one moment that IBM hasn't thought about this .... they'll want to cripple the Power5 so much that it will pose no threat to their server sales. In a way that's the beauty of buying from Intel they won't compete with you.

 

[strider42]

Originally posted by ffakr
First off... I never said the Power4 would be in a Mac. Drop that part, it's a non-issue.

Your argument doesn't make any sense though. Just because it is reliable doesn't mean it won't work in a desktop. Just because it is currently prohibitivly expensive doesn't mean it won't work in a desktop (though market forces would make it impractical).

There is nothing in the architecture of a processor that makes it inherently expensive to create. Cost is related to R&D, Demand, Supply, Production costs (size, yield,...). It isn't related to the way in which the gates are arranged in a litho mask.
Server processors do tend to have thicker gates and interconnects to increase reliabilty, but even this is a minor extra cost. They are also sorted to a higher standard than desktop parts, increasing testing cost and decreasing yeild.
Once the 'maze' of the architecture is laid down, the trivial part would be shrinking the 'walls'. It would be a LOT easier than maintaining an architecture and a derivative of that architecture.

Now, look at my argument again. What makes more sense, making a Power5 for the desktop/small server or making a Power5 and yet another architecture based off of the Power5? The cost associated with making a desktop Power5 is MUCH lower than making a derivative architecture... and the dev time would be much shorter too.

As I said... I don't know if a Power5 will end up in a Mac and you certainly don't either. IBM is hinting (over a year out) that this processor will run in smaller hardware than the Power4 was designed to run in. The Power4+ (smaller die) is being put into smaller servers already. Add to this the fact that NO ONE knows what a Power4 really costs because the only info you have is the cost of an IBM aftermarket upgrade part (which is an IBM branded module with a big L3 cache)... not a raw chip price. IBM doesn't sell them to anyone else, so we have no idea. We have no idea how increased production and increased supply for a much larger market would affect the production cost of a chip like the Power4+ and we certainly don't know how it would affect a processor like the Power5... which just got powered on 2 weeks ago.

I think the key to whether we may see a real Power5 in a Mac would be whether or not IBM includes Altivec in the Power5, and what the transistor count turns out to be.

One more thing... if IBM does design the Power5 to be an option for the Mac, the increased production will drive down per unit cost. This means larger margins and more competitive prices for their Enterprise hardware.... their largest business unit.

reliability has a direct effect on speed. The pwoer4's are made to be absolutely bullet proof. I froget allt he technical details, but its absolutely affects how fast a processor is. IBM wasn't going for the fastest chip whenh they made the power4. They were going for the best server chip, which includes being about the most reliable chip on the market. This makes it more expensive and sacrifices speed. This is absolute fact. My argument is sound.

With that in mind, thats why apple would never use it in a desktop, nor would anyone else. It makes it expensive, it makes it big, it makes compromises that don't need to be made for desktop machines. You don't use a mac truck to get around town.

It makes perfect sense for IBM to make a server chip, and then make a derivitive of it. You can't use one size fits all chips when you are making servers that costs in the hundreds of thousands of dollars range. The power4 and power5 will never be in a desktop. Is it possible the power5 will be more flexible, sure it is. But there is absolutely no evidence that it will be a desktop chip, whereas it is absoltuely positive it will be s erver chip. Given that, and the pecedent with the power4 and 970, it makes much more sense that there may be a power5 derivitive chip, but it won't be the power5.

I'm not trying to just be argumentative, but just trying to point out there is absolutely nothing anywhere that points to the power5 being anything more than a server chip, where as there is evidence and precedent for a derivitive chip to be made, adn that server and desktop chips have different priorities.

 

[strider42]

Originally posted by ffakr
I believe that Power4+ IS as fast or faster than the 970 when you don't consider the MHz increase.

from SPEC.org
IBM pSeries 650 Model 6M2 (1450 MHz, 1 CPU) 1 909 935
IBM eServer 655 Model 651 (1300 MHz, 1 CPU) 1 822 848

estimated SPEC for the 970 at 1.8 GHz are about the same for Int and slightly higher for FP. IBM says the 970 numbers will likely go up as they near production, but I think a 1.45GHz Power 4 would hold its own against a similarly clocked 970.... don't you?

Yes, but IBM has stated that the 970 would be faster than the power4 in some instances, I believe because of the priorities of design in the chip. The power4 is probably faster overall in a lot of areas, but the 970 will hold its own against it. Which is relaly the point. The 970 will be smaller, cheaper and mroe available, and over similar performance to a top tier server chip.

 

[ffakr]

Originally posted by applejilted
The danger with IBM is, IMHO, that IBM is also a a major server manufacturer and as such wants to incorporate its chips in its boxes. Apple has a modestly successful server product (Xserve with OS X server software) will IBM want to supply Apple with their Power5 chip (albeit not the multicore variety) so Apple could compete with IBM in what IBM considers a crucial market ?

Insightful point... but IBM's idea of a server and Apple's idea of a server are radically different. IBM make Billions (with a big FAT B) by selling Big Iron. That is mainframes and boxes with many CPUs that cost over $100,000 each (on the low end). Apple will never be a competetor here.
IBM has only recently started putting Power4+ chips (within the last few days) into their 'low end' unix boxes... the ones that start at $15,000 stripped... and usually ship with a much bigger price tag than that.

this could all change if Apple goes after Grid computing in a big way with PPC based clusters and small commodity linux servers, but this isn't a big component of their sales at this point.
... It's not like apple is going to release 16 or 32way PPC 970 boxes this fall.

 

[ffakr]

Originally posted by strider42
Yes, but IBM has stated that the 970 would be faster than the power4 in some instances, I believe because of the priorities of design in the chip. The power4 is probably faster overall in a lot of areas, but the 970 will hold its own against it. Which is relaly the point. The 970 will be smaller, cheaper and mroe available, and over similar performance to a top tier server chip.

The power 4 is also designed to be a dual core processor and it keeps up with the 970 with only one of if its cores. It is important to note that any Power4 (unless intentionally crippled by IBM) would post nearly twice the scores of the 970 due to the dual cores.
Of course the 970 will be smaller, it's less than half the chip, with Altivec tacked on. The 970 doesn't need to be designed with the extra dual core logic that makes the Power4 such a great processor.

 

[sergeantmudd]

Originally posted by applejilted
The danger with IBM is, IMHO, that IBM is also a a major server manufacturer and as such wants to incorporate its chips in its boxes. Apple has a modestly successful server product (Xserve with OS X server software) will IBM want to supply Apple with their Power5 chip (albeit not the multicore variety) so Apple could compete with IBM in what IBM considers a crucial market ?

I remember reading about IBM couldn't meet the demand for p690 when it came out because there was a shortage of Power4 chips. The shortage of Power4 chips was a result of their chip foundry being too busy stamping out PA-RISC and Alpha chips for HP and Compaq. IBM realizes that companies like Apple, HP, and Compaq are going to be competitive with or without IBM's help, so IBM might as well make some money helping them out.

 

[firestarter]

Originally posted by ffakr
first off... I compared IBMs HT implementation to Intels HT... I didn't comprare the 970 to the P4. You are taking my comment out of context.

The only point I wanted to make here was to point out that Intels HT is being addressed and improved.

wow, another P4 built on the same core with upgraded stats. 8K of additional L1 cache (btw, I'd expect this to be instruction not data cache.. but what do I know).
So intel continues to build ever larger processors in the pursuit of ever diminishing returns. The Intel x86 project is quicly becoming the stereotypical 'American Musclecar'. Big, powerful, and not especially efficient. Even at .09 micron, I'm sure this will be a big chip... certainly big by the standards set by other .09 micron chips.

Well spotted about the 'data cache' - you're probably right (original quote is from Anandtech - they should be more careful).

I have an engineering background and like elegant designs as much as the next person... but unfortunately this game isn't a design beauty parade, it's about IBM/Moto/Apple keeping up with Intel in speed.

I guess we'll have to wait to see if Intel runs into problems cooling such this sucker considering the number of gates in such a dense array.

Agreed. Intel seems to be wise to this - whitness the birth of Bandinas/Centrino on the laptop - where Intel's gone back to shorter pipelines and more design efficiency to build a chip that's much more power/clock efficient than the P4.

Here is where your argument falls apart. The Prescott core will scale to 4-5 GHz. That is the speed predicted over the life of the core. The 970 will likely DEBUT at 1.5-2GHz, and the Prescott will DEBUT lower than 4-5GHz, but it may eventually scale that high. This is Extremely important. We don't know but a 1.8 or 2GHz introduction of 970 on .13 micron will very likely scale north of 3GHz on .09 micron. Will a 4GHz Prescott be faster than a 3GHz 970? We have no idea. The current SPEC seems to indicate that the 970 would give it a damn fine run for its money.

Actually I disagree. SPECmark ratings have been widely published - one report was on:
http://www.aceshardware.com/read_news.jsp?id=60000475

This shows the 1.8G 970 testing slower than the 3.06G P4. I would expect an improved P4 core with cache/architectural improvements, running at 4GHz to be faster!

Another thing to consider... do you expect a 5GHz P4 core to be 25% faster than a 4GHz P4 core? You don't expect these ultra high clock speeds to show substantially diminishing returns as the CPU runs away from the speed of the components?

Front side bus speed is also massively increased, and the L2 cache size is doubled. Returns will deminish, but I would expect most of this increase in speed to be realised. On the desktop, media handling (compression/decompression etc.) is the application most folks will want a speedup in, and this is tight code which will fit in the cache.

IMHO... the key to this decade isn't 10GHz processors, it's more intelligent design. IBM is on the right track with what appears to be a much better SMT implementation and a smaller, more efficient cpu core. Even Sun seems to get it. They expect to have massive SMT and multi-cores per die... Sun hopes to have 8, 16, maybe 32 threads per processor die by 2005.

I agree, except that I'd add that the important thing will be software design. Most folks want to see a speed increase in the straight line performance of the single app they're running, and this is an important difference in requirements to the Sun server chips you cite. Apple need to get their developers making better use of threads and Altivec, so that users do experience these speed increases on parallel processors.

 

[type_r503]

quote from page 2

"Robert Scoble writes in his weblog:

I've had some sneaks behind the scenes (not official ones, though). Apple has some cool stuff coming this year to be sure -- including some desktop machines that are outperforming current Intel stuff."