<strong>1) It might get sticky trying to share one SIMD engine, so do we need two of these also on one chip?</strong><hr></blockquote>
There were some images of a Power4 posted some time ago, the cip layout was basically symmertrical around the centre-line of the chip - so'd I'd pretty much expect a "dual core" to double-up on available units.
Thanks to all for the inputs. I can understand two complete 970 processors on one chip. I figured that sharing things would be a nightmare, but I wasn't sure. Going from 130 nm to 90 nm means the area is cut about in half. So a two core processor would yield the same number of chips per wafer on the smaller process.
I never considered higher clock speed as an alternative, nor making a larger, more complex core, but I was comparing a two core chip with having two single core processors, the dual processors scheme we have today. I believe heat may still be the biggest concern. What I don't know is how heat scales with process size. If the chip area is cut in half, I have a feeling that power dissipation is not cut in half, at the same clock speed. If someone has the facts on this it would be interesting. If heat does scale the same as area, it would be cool. (Oops, that pun was not intended.) It means there can be a two core package with the same heat and chips per wafer, as the present 130 nm process. However, clock speed still enters the picture.
Let's assume we have a 970 processor that has a certain clock speed, which is determined to keep power dissipation within manageable limits. With a smaller process, clock speed can be raised and the processor will still run at this same power level. I see where the trade off may be a two core processor package running at a lower clock speed, or two single core processor packages running at higher clock speeds. The dual core will have better communication between the the two cores, but the two single core processors are each running faster.
If everybody is planning for multi-core packages, there is obviously more to consider. Possibly, heat dissipation for the package is not a limiting factor at all.
<strong>If everybody is planning for multi-core packages, there is obviously more to consider.</strong><hr></blockquote>
One of the aspects is the purely physical layout bits. The bit of silicon that is "the CPU" has to have hundreds of wire traces hooked up to it, and slammed into a package. The ppc970 (for instance) has 576 wires, 161 signal wires.
If we have two separate CPUs, we have (exactly) twice as many traces coming out of twice as many packages, and we also increased the complexity of the next chip in line, call that the "northbridge". That chip _also_ has to have a lot more wires. One set has to go to CPU1, and one set to CPU2... and all of this needs to be designed/tested to make sure that it can run full tilt without causing radio interference with all the other signals running around.
If we have one CPU with 2 cores, there is a little bit more than double the number of transistors inside - the extra transistors mean that we don't have to go to 1152 pins for our dual-core 970 chip. Nor do we need two separate sets of traces between the CPU and the northbridge. The part that the two cores share is the access to the outside world. All that says is there is some point where it would be cheaper to build a CPU that is twice as complex (dual core) to avoid paying for all the extra wires/chips etc or supporting two physical CPUs.
Or you could think of it this way: 8bit CPUs can be ganged together into truly gargantuan computers that exceed the output of current Macs. Why aren't we doing that? -> It isn't because the 8bit CPUs are expensive, it is all the wiring & design work running all over the place that would make that impractical. Yet the Altivec unit can pretend to be a pile of 8/16/32 bit computers all running the same thing - and do so cheaply.
. . . If we have two separate CPUs, we have (exactly) twice as many traces coming out of twice as many packages, and we also increased the complexity of the next chip in line, call that the "northbridge". . .
</strong><hr></blockquote>
That's true. It looks like there is more overhead to a dual processor computer than most of us realize. (Makes me wonder how much extra cost there is a PowerMac now that they are all duals.) Yet, there has been so much discussion and excitement about the ability of the 970 to easily work in duals, quads and even higher. It looks to me like it may be a matter of economics. A dual core processor package would cost less. However, we could still have a trade off between a slower, less costly dual core package, and two higher clock speed single core packages, which cost more but may have better performance. (This assumes that power is a limiting factor, as I suggested in my previous post.)
<strong>However, we could still have a trade off between a slower, less costly dual core package, and two higher clock speed single core packages, which cost more but may have better performance. (This assumes that power is a limiting factor, as I suggested in my previous post.)</strong><hr></blockquote>
And I'm hoping the trade off would be between one dual core 970's... and TWO. You are right, heat dissipation is very important.
I don't know that we'll see something like this:
1.5 GHz dual core ppc 9x0
2.0 GHz single core ppc 970
at the same time. If they can make a 2.0 GHz single core with a reasonable success rate, they can try to make dual cores at that same time. -> any time one of the cores fails they say 'oops, looks like a single core module, throw it in the 2.0 GHz single core ppc box'. But every time it _works_ they have a dual core.
That is, it is expensive to have widely different chip fab lines running at the same time -> they'd all come off of one line. The 'defects' are sold as either lower GHz parts, or single core parts or whatever - but they shouldn't be drastically different. And the dual core parts would always cost more -> tougher to make. (Any time a dual core 'fails' you get a "free" single core)
I have no idea how the heat trade off will end up working though.
Yea but how is Apple going to market dual core CPU's? You can't say dual 1.8ghz or can you? Maybe we'll see dual dual core CPU's...wow! I guess what I'm getting at is if the 970 became dual core would apple still put two of them in their Powermacs? Or would they only put one dual core chip and call it a dual system. To be dual or not to be...
<strong>Yea but how is Apple going to market dual core CPU's? You can't say dual 1.8ghz or can you? Maybe we'll see dual dual core CPU's...wow! I guess what I'm getting at is if the 970 became dual core would apple still put two of them in their Powermacs? Or would they only put one dual core chip and call it a dual system. To be dual or not to be...</strong><hr></blockquote>
Well, They may market a dual core as a "dual" and finally get around to doing a dual "dual-core" and market it as a quad. After all, IBM has been testing uni, dual and quad configurations for some time now. And yes, OSX has been tested with core apps (not to be confused with processor cores) on the first two cases. Unfortunately, we gonna have to wait until the 90nm process is implemented. But just think, when dual-cores do come out, the consumer end of the mac line may get the single core "washouts" of the 970 line.
. . . (Any time a dual core 'fails' you get a "free" single core)
I have no idea how the heat trade off will end up working though.
</strong><hr></blockquote>
I believe they would cut the power supply to the bad core when they package it as a single core. In that way, it could run at a higher clock rate without exceeding the power limit, or just run cooler at the same clock rate. I'm not sure whether this is what you are also saying.
I don't think there would be enough dual-core failures to create enough washout chips to be used in the iMacs. I would feel funny using a computer a half working CPU. Any way all this dual core stuff is still a long way off. For now we should be more concerned with whether apple will keep the PowerMacs all dual when the 970 comes out.
<strong>I don't think there would be enough dual-core failures to create enough washout chips to be used in the iMacs. I would feel funny using a computer a half working CPU. Any way all this dual core stuff is still a long way off. For now we should be more concerned with whether apple will keep the PowerMacs all dual when the 970 comes out.</strong><hr></blockquote>
Valid points. Maybe enough to sell low priced single processor towers though. Another bad side to it would be the road to single supplier. As it would drive Moto out. Granted, I don't care for Moto's past history, but being limited to a single supplier is worse.
As for Apple keeping all duals ... It's anybody's guess. Mine being that apple will go back to the mixed set (uni and dual models).
<strong>... I would feel funny using a computer a half working CPU. .</strong><hr></blockquote>
Actually, this is what IBM is doing now with the Power 4 chips. Dice that only have one core working are put into lower-end servers instead of the top-of-the-line.
Also, I don't believe that the 'one working core' chips are all that common. It is more common for either both cores to function, or none at all.
Yea but dual-core sounds better! Imagine the apple adds..." you get 4 CPU's for the price of two!" lol... Or if they cheap us out of the dual configuration they can still claim that the dual core is like having two CPU's. Which I'm thinking it is.
<strong>Would it be more practical to have a hyper-threaded single core with twice the execution units?</strong><hr></blockquote>
The HyperThreading is supposed to be in the Power5 -> the generation after this. So then you'd have dual dual cores pretending to be dual quad cores via HyperThreading. I can see why Apple is harping on the 'thread everything' mantra for the developers
It isn't something IBM can adopt quickly. They designed the Power4 quite some time ago now - adding a block like AV is a lot easier than rewiring all of the internals for HyperThreading.
I guess the only thing I wonder about wih this rumors is the part about Marklar.
The idea of mass user disatisfaction and defections to other platforms due to the actual implementation of Microsoft and Intel's DRM/spyware efforts isn't new in discussion about DRM. I think it's smart for Apple to want to take advantage of the situation if it occurs. This kind of stuff has just been experimented with so far (A few CD's here and there) but fully implemented Apple could see a significant chance to increase market share.
However, previous discussions about OSX on x86 has centered around Apple's difficulty with chip suppliers thus forcing them to change their hardware to x86 (or perhaps x86 64). These machines would likely be proprietary machines.
This rumor however says nothing about that, instead it seems to suggest that in the event that you had millions of pissed off XP users so frustrated with getting anything done on their machines in privacy that they were ready to switch, Apple would offer them a shiny new OSX CD that they could install on their current PC.
My questions is this. Would Apple really give up control of what hardware OSX is run on and throw themselves back into the "Clone Wars" again just to gain that market share or would they try to get these people to switch to Apple PPC hardware? Some would probably go to Linux rathar than buy a new PC I'm sure but as we all know running an OS on a platform where everyone and their brother builds machines with a hundreds of different possible hardware combinations (chip sets, video cards, network cards, etc) is a support nightmare. Some of the problems with PC's are certainly the fault of Windows but not all. This is where the Macintosh gets it's low total cost of ownership.
This rumor isn't far fetched considering the reaction to Microsoft licensing recently. I have no doubt what has been outlined in this rumor has been discussed as a contingency but I'm not so sure Apple would actually do it.
Another reason for having an x86 version of OS X could be to run comparisons. Apple can compare how their own applications run on different processors. Even if they never change over, it shows them where more effort needs to be placed in PPC improvements.
<strong>Would it be more practical to have a hyper-threaded single core with twice the execution units?</strong><hr></blockquote>
The problem with a hyper-threaded core is that resources are shared so that the threads need to be in the same process, and those threads should be doing different things (i.e. integer/float, float/vector, integer/vector). Also, if a manufacturing flaw kills part of a hyper-threaded processor then the whole processor is dead.
Single processor failures in a dual processor chip are common enough that IBM is shipping whole lines of machines based on them! Even if it is a fairly low percentage, this practice is only going to become more common as the number of cores grows. If you have 16 cores and a flaw kills one (or two, or three, etc) of them but your design still allows the others to operate independently, then why not sell it as a less powerful chip? This is going to be as important in the future as selling different chips speeds is today.
Comments
<strong>1) It might get sticky trying to share one SIMD engine, so do we need two of these also on one chip?</strong><hr></blockquote>
There were some images of a Power4 posted some time ago, the cip layout was basically symmertrical around the centre-line of the chip - so'd I'd pretty much expect a "dual core" to double-up on available units.
<strong>...not to belabor it...but you Montana city slickers don't know nothing 'bout farmin' 'parently...
quote:
If our hero the 970 kicks in and does as well as it should, Apple/IBM may not have that long of a road to hoe.
....I bleeve round here, we do all our hoeing on rows.
Sorry— I’m from the city, and around here, all of our Ho's live in skid rows.
I never considered higher clock speed as an alternative, nor making a larger, more complex core, but I was comparing a two core chip with having two single core processors, the dual processors scheme we have today. I believe heat may still be the biggest concern. What I don't know is how heat scales with process size. If the chip area is cut in half, I have a feeling that power dissipation is not cut in half, at the same clock speed. If someone has the facts on this it would be interesting. If heat does scale the same as area, it would be cool. (Oops, that pun was not intended.) It means there can be a two core package with the same heat and chips per wafer, as the present 130 nm process. However, clock speed still enters the picture.
Let's assume we have a 970 processor that has a certain clock speed, which is determined to keep power dissipation within manageable limits. With a smaller process, clock speed can be raised and the processor will still run at this same power level. I see where the trade off may be a two core processor package running at a lower clock speed, or two single core processor packages running at higher clock speeds. The dual core will have better communication between the the two cores, but the two single core processors are each running faster.
If everybody is planning for multi-core packages, there is obviously more to consider. Possibly, heat dissipation for the package is not a limiting factor at all.
<strong>If everybody is planning for multi-core packages, there is obviously more to consider.</strong><hr></blockquote>
One of the aspects is the purely physical layout bits. The bit of silicon that is "the CPU" has to have hundreds of wire traces hooked up to it, and slammed into a package. The ppc970 (for instance) has 576 wires, 161 signal wires.
If we have two separate CPUs, we have (exactly) twice as many traces coming out of twice as many packages, and we also increased the complexity of the next chip in line, call that the "northbridge". That chip _also_ has to have a lot more wires. One set has to go to CPU1, and one set to CPU2... and all of this needs to be designed/tested to make sure that it can run full tilt without causing radio interference with all the other signals running around.
If we have one CPU with 2 cores, there is a little bit more than double the number of transistors inside - the extra transistors mean that we don't have to go to 1152 pins for our dual-core 970 chip. Nor do we need two separate sets of traces between the CPU and the northbridge. The part that the two cores share is the access to the outside world. All that says is there is some point where it would be cheaper to build a CPU that is twice as complex (dual core) to avoid paying for all the extra wires/chips etc or supporting two physical CPUs.
Or you could think of it this way: 8bit CPUs can be ganged together into truly gargantuan computers that exceed the output of current Macs. Why aren't we doing that? -> It isn't because the 8bit CPUs are expensive, it is all the wiring & design work running all over the place that would make that impractical. Yet the Altivec unit can pretend to be a pile of 8/16/32 bit computers all running the same thing - and do so cheaply.
<strong>
. . . If we have two separate CPUs, we have (exactly) twice as many traces coming out of twice as many packages, and we also increased the complexity of the next chip in line, call that the "northbridge". . .
</strong><hr></blockquote>
That's true. It looks like there is more overhead to a dual processor computer than most of us realize. (Makes me wonder how much extra cost there is a PowerMac now that they are all duals.) Yet, there has been so much discussion and excitement about the ability of the 970 to easily work in duals, quads and even higher. It looks to me like it may be a matter of economics. A dual core processor package would cost less. However, we could still have a trade off between a slower, less costly dual core package, and two higher clock speed single core packages, which cost more but may have better performance. (This assumes that power is a limiting factor, as I suggested in my previous post.)
<strong>However, we could still have a trade off between a slower, less costly dual core package, and two higher clock speed single core packages, which cost more but may have better performance. (This assumes that power is a limiting factor, as I suggested in my previous post.)</strong><hr></blockquote>
And I'm hoping the trade off would be between one dual core 970's... and TWO.
I don't know that we'll see something like this:
1.5 GHz dual core ppc 9x0
2.0 GHz single core ppc 970
at the same time. If they can make a 2.0 GHz single core with a reasonable success rate, they can try to make dual cores at that same time. -> any time one of the cores fails they say 'oops, looks like a single core module, throw it in the 2.0 GHz single core ppc box'. But every time it _works_ they have a dual core.
That is, it is expensive to have widely different chip fab lines running at the same time -> they'd all come off of one line. The 'defects' are sold as either lower GHz parts, or single core parts or whatever - but they shouldn't be drastically different. And the dual core parts would always cost more -> tougher to make. (Any time a dual core 'fails' you get a "free" single core)
I have no idea how the heat trade off will end up working though.
<strong>Yea but how is Apple going to market dual core CPU's? You can't say dual 1.8ghz or can you? Maybe we'll see dual dual core CPU's...wow! I guess what I'm getting at is if the 970 became dual core would apple still put two of them in their Powermacs? Or would they only put one dual core chip and call it a dual system. To be dual or not to be...</strong><hr></blockquote>
Well, They may market a dual core as a "dual" and finally get around to doing a dual "dual-core" and market it as a quad. After all, IBM has been testing uni, dual and quad configurations for some time now. And yes, OSX has been tested with core apps (not to be confused with processor cores) on the first two cases. Unfortunately, we gonna have to wait until the 90nm process is implemented. But just think, when dual-cores do come out, the consumer end of the mac line may get the single core "washouts" of the 970 line.
<strong>
. . . (Any time a dual core 'fails' you get a "free" single core)
I have no idea how the heat trade off will end up working though.
</strong><hr></blockquote>
I believe they would cut the power supply to the bad core when they package it as a single core. In that way, it could run at a higher clock rate without exceeding the power limit, or just run cooler at the same clock rate. I'm not sure whether this is what you are also saying.
<strong>I don't think there would be enough dual-core failures to create enough washout chips to be used in the iMacs. I would feel funny using a computer a half working CPU. Any way all this dual core stuff is still a long way off. For now we should be more concerned with whether apple will keep the PowerMacs all dual when the 970 comes out.</strong><hr></blockquote>
Valid points. Maybe enough to sell low priced single processor towers though. Another bad side to it would be the road to single supplier. As it would drive Moto out. Granted, I don't care for Moto's past history, but being limited to a single supplier is worse.
As for Apple keeping all duals ... It's anybody's guess. Mine being that apple will go back to the mixed set (uni and dual models).
<strong>... I would feel funny using a computer a half working CPU. .</strong><hr></blockquote>
Actually, this is what IBM is doing now with the Power 4 chips. Dice that only have one core working are put into lower-end servers instead of the top-of-the-line.
Also, I don't believe that the 'one working core' chips are all that common. It is more common for either both cores to function, or none at all.
<strong>Would it be more practical to have a hyper-threaded single core with twice the execution units?</strong><hr></blockquote>
The HyperThreading is supposed to be in the Power5 -> the generation after this. So then you'd have dual dual cores pretending to be dual quad cores via HyperThreading. I can see why Apple is harping on the 'thread everything' mantra for the developers
It isn't something IBM can adopt quickly. They designed the Power4 quite some time ago now - adding a block like AV is a lot easier than rewiring all of the internals for HyperThreading.
The idea of mass user disatisfaction and defections to other platforms due to the actual implementation of Microsoft and Intel's DRM/spyware efforts isn't new in discussion about DRM. I think it's smart for Apple to want to take advantage of the situation if it occurs. This kind of stuff has just been experimented with so far (A few CD's here and there) but fully implemented Apple could see a significant chance to increase market share.
However, previous discussions about OSX on x86 has centered around Apple's difficulty with chip suppliers thus forcing them to change their hardware to x86 (or perhaps x86 64). These machines would likely be proprietary machines.
This rumor however says nothing about that, instead it seems to suggest that in the event that you had millions of pissed off XP users so frustrated with getting anything done on their machines in privacy that they were ready to switch, Apple would offer them a shiny new OSX CD that they could install on their current PC.
My questions is this. Would Apple really give up control of what hardware OSX is run on and throw themselves back into the "Clone Wars" again just to gain that market share or would they try to get these people to switch to Apple PPC hardware? Some would probably go to Linux rathar than buy a new PC I'm sure but as we all know running an OS on a platform where everyone and their brother builds machines with a hundreds of different possible hardware combinations (chip sets, video cards, network cards, etc) is a support nightmare. Some of the problems with PC's are certainly the fault of Windows but not all. This is where the Macintosh gets it's low total cost of ownership.
This rumor isn't far fetched considering the reaction to Microsoft licensing recently. I have no doubt what has been outlined in this rumor has been discussed as a contingency but I'm not so sure Apple would actually do it.
[ 12-01-2002: Message edited by: nebcon65 ]</p>
<strong>Would it be more practical to have a hyper-threaded single core with twice the execution units?</strong><hr></blockquote>
The problem with a hyper-threaded core is that resources are shared so that the threads need to be in the same process, and those threads should be doing different things (i.e. integer/float, float/vector, integer/vector). Also, if a manufacturing flaw kills part of a hyper-threaded processor then the whole processor is dead.
Single processor failures in a dual processor chip are common enough that IBM is shipping whole lines of machines based on them! Even if it is a fairly low percentage, this practice is only going to become more common as the number of cores grows. If you have 16 cores and a flaw kills one (or two, or three, etc) of them but your design still allows the others to operate independently, then why not sell it as a less powerful chip? This is going to be as important in the future as selling different chips speeds is today.