[Simon]

ThinkSecret claims they have news about IBM's 970 development.

Currently an 'Antares' dubbed dual-core version of the 970 is in the works, the 970MP. It will run at 3GHz, have 1MB L2 cache and no L3. It is however not known when it will start shipping.

Its development has also spawned a single-core version dubbed 970GX or 'Antares SP' also running at 3GHz with 1MB L2. The nice thing about the GX is that it could ship as early as Q2 2005.

They also mention that nothing is really known about the status of the low-power mobile 970. At least they claim it's being tested for real at 1.6-1.8GHz.

[UnixMac]

Originally posted by Simon:
AppleInsider claims they have news about IBM's 970 development.

Currently an 'Antares' dubbed dual-core version of the 970 is in the works, the 970MP. It will run at 3GHz, have 1MB L2 cache and no L3. It is however not known when it will start shipping.

Its development has also spawned a single-core version dubbed 970GX or 'Antares SP' also running at 3GHz with 1MB L2. The nice thing about the GX is that it could ship as early as Q2 2005.

They also mention that nothing is really known about the status of the low-power mobile 970. At least they claim it's being tested for real at 1.6-1.8GHz.

The big question is, will Apple have a top end machine that is a dual "dual core", or will they revert to single CPU (dual core) models.

I would like to see single for the lower end, and possible 1 dual "dual core" at the top end for under $3500.

[ReggieX]

Eh, I'll believe it when I see it on the Apple Store. I'm in the market for a new Mac sometime in the new year, so I'll see what happens.

[Simon]

Originally posted by UnixMac:
The big question is, will Apple have a top end machine that is a dual "dual core", or will they revert to single CPU (dual core) models.

I would like to see single for the lower end, and possible 1 dual "dual core" at the top end for under $3500.

I guess dual-core won't have much influence on single or dual CPU models.

Dual-core is the processor industry's way to deal with hitting the MHz ceiling (90nm designs required too much dissipation per mm^2 at higher clock rates). It is no replacement for dual CPU systems per se.

Of course Apple could re-market the iMac as the 970SP G5 Mac and the PowerMac as 970MP, but would that sell well? Certainly not as well as single and dual CPU.

[Big Mac]

Yeah, I too am doubtful. It seems to me that TS's track record has been declining as of late, although I may be unfairly maligning them. With Apple having so much difficulty with 2.5GHz single cores, I don't see how IBM could deliver dual core 3GHz parts unless they've experienced some sort of breakthrough recently.

[iohead]

Originally posted by Simon:
I guess dual-core won't have much influence on single or dual CPU models.

Dual-core is the processor industry's way to deal with hitting the MHz ceiling (90nm designs required too much dissipation per mm^2 at higher clock rates). It is no replacement for dual CPU systems per se.

Of course Apple could re-market the iMac as the 970SP G5 Mac and the PowerMac as 970MP, but would that sell well? Certainly not as well as single and dual CPU.

You might be confusing hyperthreading with dual-core. They are not the same! Some more relevant details are in this thread:

kernelthread.com thread

Dual-core is as good as having 2 separate cpus from the point of view of smp.

[Simon]

Originally posted by iohead:
You might be confusing hyperthreading with dual-core.

No, I'm not confusing them. But from a marketing POV (which is what I was talking about) it's much better to advertise two G5s than one. Normal customers don't know jack about CPU cores and threading. They want high GHz numbers and as many CPUs as their money will buy.

That said, I think it's easier to market two 3GHz chips with two big brazed aluminum heat sink covers (with a big fat 'G5' written on them) than to explain to a customer that they'll get a single 3GHz chip that actually contains two chips on its die.

And btw, hyperthreading is something we Mac users don't really have to care about anyway. Hyperthreading is being promoted big time by Intel in order to overcome limitations of their antique P4 x86 design that doesn't do SMP. Outside of the server world nobody really uses Xeons (which can do SMP) and therefore to do something like SMP in the broader x86 world you need Hyperthreading. We have real SMP CPUs right away, so we could care less about it as long as we have either dual 970s or single 970MPs.

[iohead]

Originally posted by Simon:
No, I'm not confusing them. But from a marketing POV (which is what I was talking about) it's much better to advertise two G5s than one. Normal customers don't know jack about CPU cores and threading. They want high GHz numbers and as many CPUs as their money will buy.

Well, I asked if you are confusing them because you stated the following:

Dual-core is the processor industry's way to deal with hitting the MHz ceiling (90nm designs required too much dissipation per mm^2 at higher clock rates). It is no replacement for dual CPU systems per se.

The definition of what a "G5" means is up to Apple. What's to prevent Apple from saying (marketing-wise) that a dual-core chip is "two processors"? It's technically correct even!

Normal customers don't know about cores, _exactly_, so why should they care how many physical chips constitute their computer? If a single "chip" gives them two processors, Apple can market it as two processors.

That said, I think it's easier to market two 3GHz chips with two big brazed aluminum heat sink covers (with a big fat 'G5' written on them) than to explain to a customer that they'll get a single 3GHz chip that actually contains two chips on its die

You seem to be implying that a "chip" is the same as a "processor". A dual-core "chip" has two *processors* on board, each running at the same clock frequency. So a "single 3 GHz chip" is not a bad thing or such: that's just an implementation detail! Each "processor" on the "chip" will be running at 3 GHz in that case. As for the heat sink cover, well, it could have "G5" etched twice.

I don't see how this is a marketing problem, or why the customer would be bothered by this "detail". As long as nobody is cheating anybody, that is, two processors means two processors, why should anyone care about the number of physical chips they are getting?

[Eug Wanker]

Nitpick: It's TS not AI.

Originally posted by Simon:
Dual-core is the processor industry's way to deal with hitting the MHz ceiling (90nm designs required too much dissipation per mm^2 at higher clock rates). It is no replacement for dual CPU systems per se.

Yes it is.

Of course Apple could re-market the iMac as the 970SP G5 Mac and the PowerMac as 970MP, but would that sell well? Certainly not as well as single and dual CPU.

Apple could sell a 970MP as a dual, and they wouldn't be lying. It would be interesting to see a $4000 quad (dual-dual) too.

Anyways, while the dual-core is interesting, I just as interesting is the 1 MB L2, and more interesting is the low power G5 for the PowerBooks.

I'll believe it when I see it... BTW, it's just two months until Macworld. (I don't expect this stuff at Macworld. Maybe at WWDC though.)

[Simon]

What you two don't understand is that customers like to see two chips. They don't care how many cores are on there. There could be four cores on a single die and still people rather see two chips regardless of the number of cores on there. For most people chip = processor = core = power.

You can ask for dual cores as much as you want, but unless the dual core is faster, cheaper and offers vastly more power, people will prefer dual single-core chips. Wanna bet?

[Simon]

Originally posted by Eug Wanker:
Nitpick: It's TS not AI.

Yep. Stupid mistake. I'm always mixing them up. Corrected.

Yes it is.

Why should it? You can do SMP regardless of the number of cores. It's not the same ball park. Just because the resulting number of cores is equal does not mean it's the same if you do it on one chip (multiple cores) or with several single-core CPUs (SMP).

Multiple cores and SMP can coexist. It's not as if one would replace the other. Actually, they won't. It will probably always be easier to add additional chips to a board than add additional cores to a die.

Apple could sell a 970MP as a dual, and they wouldn't be lying. It would be interesting to see a $4000 quad (dual-dual) too.

Of course they can. The question is if people will go for it. I claim people prefer two single-core chips to one dual-core chip. Marketing will have to overcome that barrier.
From a technical POV multi-coring is probably the way to go. Much less overhead needed. But that's another issue.

[Eug Wanker]

Originally posted by Simon:
What you two don't understand is that customers like to see two chips. They don't care how many cores are on there. There could be four cores on a single die and still people rather see two chips regardless of the number of cores on there. For most people chip = processor = core = power.

You can ask for dual cores as much as you want, but unless the dual core is faster, cheaper and offers vastly more power, people will prefer dual single-core chips. Wanna bet?

Quite frankly, I think you're barking up the wrong tree.

I think most customers really couldn't care less if it's two single cores or one dual-core. In fact, I betcha 90% of Apple customers don't even know what that menas. As long as a dual-core single chip 3.0 is faster than a dual single-core dual chip 2.5, people will buy it. That's all that really matters after all.

Also, with the dual 2.5 you can't physically see ANY chips anyway. It's all covered by a liquid cooling system and cover. You can "see" the chips via software, but you'd see the separate cores on a dual-core system too.

I really think all of this argument about dual-core marketing is really irrelevant. Besides, even if it were relevant, Apple has one of the best marketing divisions in existence. More important is whether or not IBM can actually make these chips at high clock speeds.

And why would it matter? Because dual-core is a theoretically cheaper way of making duals. (Mind you it could just be that if IBM did make these chips and sold them to Apple for cheaper than two single-core chips, Apple could initially just pocket the difference instead of passing on the savings to consumers.)

[Simon]

Originally posted by Eug Wanker:
As long as a dual-core single chip dual 3.0 is faster than a dual 2.5, people will buy it. That's all that really mattes after all.

Certainly.

I really think all of this argument about dual-core marketing is really irrelevant. Besides, Apple has one of the best marketing divisions in existence.

The argument comes from the fact that it was already suggested in this thread, Apple could drop the duals across the line and go with the GX for the lower end and the MP for the higher end. And I claim from a marketing POV that sucks and if it's just that now people see two big mucho-macho heat sinks and in the future there will be just one. Granted, that's ridiculous, but it's exactly the kind of talk going on in Apple stores. Go listen.

It would however be grand to go dual MP on the top. 4 cores in one system, magnificent. That would be very expensive, but offer the full glory money can buy.

Because dual-core is a theoretically cheaper way of making duals.

Bingo. And that's why it's a good thing. But nobody should suggest duals have to die once we get dual cores. Actually, dual cores and dual CPUs together makes me want to spend money big time.

[Eug Wanker]

Originally posted by Simon:
Why should it? You can do SMP regardless of the number of cores. It's not the same ball park. Just because the resulting number of cores is equal does not mean it's the same if you do it on one chip (multiple cores) or with several single-core CPUs (SMP).

Multiple cores and SMP can coexist. It's not as if one would replace the other. Actually, they won't. It will probably always be easier to add additional chips to a board than add additional cores to a die.

I'm not sure what you're talking about. I'm no expert myself, but it sounds like you don't know what SMP means in this context. The OS sees the cores as separate regardless of whether or not they're on the same chip. SMP does NOT mean the cores must be on separate chips. Conceptually, SMP on dual-core is basically the same thing as SMP on dual single-core.

ie. With an OS that can handle SMP:

2 chips with 1 core each = 2 CPUs that can be SMP'd.
1 chip with dual cores = 2 CPUs that can be SMP'd.

And it's architecturally simpler to have multiple cores on a single chip than it is have multiple discrete CPUs.

[Simon]

Originally posted by Eug Wanker:
I'm not sure what you're talking about.

The OS sees the cores as separate regardless of whether or not they're on the same chip.

SMP does NOT mean the cores must be on separate chips.

AHHHHH. I know.

OK. really slowly. If you take a 970 you get one core. You can take two 970s and you get two cores (SMP). Nice.

If you now take a 970MP you get two cores (SMP as well). And if you take 2 970MPs you get four (also SMP). Even nicer.

But now, suppose I want 8 cores. What do I do? Wait for IBM to develop the fire-sucking god SUPER970MP that has 8 cores or should I just take 4 970MPs and put them on one board? Which is easier? Exactly. And that's why multiple CPU boards and multi-cores do not exclude each other. Of course once you have dual cores it's cheaper than two single cores, but the problem is that manufacturers go through great trouble to get multi-cores whereas putting several available chips on a board is rather simple. Add to that that the cost will grow non-linearly with the number of cores. Just because we'll have dual cores in a year, does not mean the number of cores will continually grow. It's not exactly trivial to accomplish.

[UnixMac]

Originally posted by Eug Wanker:
I'm not sure what you're talking about. I'm no expert myself, but it sounds like you don't know what SMP means in this context. The OS sees the cores as separate regardless of whether or not they're on the same chip. SMP does NOT mean the cores must be on separate chips. Conceptually, SMP on dual-core is basically the same thing as SMP on dual single-core.

ie. With an OS that can handle SMP:

2 chips with 1 core each = 2 CPUs that can be SMP'd.
1 chip with dual cores = 2 CPUs that can be SMP'd.

And it's architecturally simpler to have multiple cores on a single chip than it is have multiple discrete CPUs.

What about the memory bus and other communications to the core.. is there an advantage to the current setup with two separate CPU's vs a dual core in this respect?

[Eug Wanker]

Originally posted by Simon:
AHHHHH.

OK. really slowly. If you take a 970 you get one core. You can take two 970s and you get two cores. Nice.

If you now take a 970MP you get two cores. And if you take 2 970MPs you get four. Even nicer.

But now, suppose I want 8 cores. What do I do? Wait for IBM to develop the fire-sucking god SUPER970MP that has 8 cores or should I just take 4 970MPs and put them on one board? Which is easier? Exactly. And that's why SMP and multi-cores do not exclude each other. Of course once you have dual cores it's cheaper than two single cores, but the problem is that manufacturers go through great trouble top get multi-cores whereas putting several available chips on a board is rather simple. Add to that that the cost will grow non-linear with the number of cores. Just because we dual cores in a year, does not mean the number of cores will continually grow. It's not exactly trivial to accomplish.

An 8 CPU Power Mac would be pointless in the near-term, whether it be 4 970 MPs or 8 single-CPUs or 1 massive 8-core chip. It would be extremely expensive, incredibly difficult to design, and it would probably have questionable potential performance anyway because of other bottlenecks in the system. Indeed, at 4 CPUs, things start becoming a tradeoff.

And what does this have to do with dual-core 970MPs coming out? It's the logical progression. Nobody is suggesting 8 CPU systems at this point. In fact, I think 4 CPU systems are unlikely in the near-term, even if the 970MP becomes available. The market for those types of systems is very small. However, if we do see 4 CPU systems, you can damn well be sure they're going to be 2 x dual-core, not 4 x single-core, cuz dual-dual is easier.

Originally posted by UnixMac:
What about the memory bus and other communications to the core.. is there an advantage to the current setup with two separate CPU's vs a dual core in this respect?

I'm not an engineer, but...

Assuming they don't add a memory controller to the chip, yes, they'd be sharing a bus, but OTOH, they'd have uber fast interchip communication, and they could also share the aggregate 2 MB L2 cache.

The Power Macs' memory bus and system bus are pretty fast, but would become bigger bottlenecks as number of cores increases. At two cores sharing a bus it's not a deal-breaker for most stuff though.

We can pretty much ignore anything about 4-core chip designs, because that ain't gonna happen any time soon in Macs. We can also ignore anything about 4 x dual-core Macs too, since that ain't gonna happen any time soon either.

[Simon]

I don't know what the is for. I'm just trying to explain the point I was making, obviously my wording was unclear.

Originally posted by Eug Wanker:
However, if we do see 4 CPU systems, you can damn well be sure they're going to be 2 x dual-core, not 4 x single-core, cuz dual-dual is easier.

Of course, once you have dual cores it's easier. But it's an enormous accomplishment to get there. And getting up to four cores will be even harder. And that's why stacking CPUs next to each other will remain a 'simple' and cheap way to do more SMP. Which brings us back to my initial point: Multi-cores and multiple CPU boards do not exclude each other.

[Simon]

Originally posted by Eug Wanker:
We can pretty much ignore anything about 4-core chip designs, because that ain't gonna happen any time soon in Macs. We can also ignore anything about 4 x dual-core Macs too, since that ain't gonna happen any time soon either.

Bill Gates ignored that 640k RAM wasn't always going to be enough.

We are in the process of buying dozens of quad Xeon boxes for our lab. Quad cores is not tabu.

Saying we Mac users should not care about it is saying we only need Fisher Price computers.

[Eug Wanker]

Originally posted by Simon:
I don't know what the is for. I'm just trying to explain the point I was making, obviously my wording was unclear.

Of course, once you have dual cores it's easier. But it's an enormous accomplishment to get there. And getting up to four cores will be even harder. And that's why stacking CPUs next to each other will remain a 'simple' and cheap way to do more SMP.

Actually, from what I've read, making the 970 a dual-core chip shouldn't be that big of a deal in the greater scheme of things. Remember, the 970 is basically one of the two cores of the POWER4, with much of the support stuff removed. ie. The 970 BEGAN LIFE as dual-core, and then was stripped of everything into a nice light single-core unit. If these rumours are true, IBM is now going back and reintegrating it into dual core, but again without all the support stuff that the POWER4 has. IOW, the 970 is the POWER4-ultralite, and the 970MP would be the POWER4-lite.

Anyways, it would be a major pain to design desktop machines to support 4 discrete CPUs, at least if you're talking 970-class chips. Such a motherboard would be extremely costly, esp. if you wanted good performance. If you're talking future directions, I would think radical design changes are in order if clock speeds can't be increased, instead of just slapping together more CPUs onto existing designs.

Which brings us back to my initial point: Multi-cores and multiple CPU boards do not exclude each other.

Actually, in a way it does. If Apple decides it's going to sell no more CPUs than 2 in a Power Mac, and the 970MP did exist in sufficient quantities and was reasonably priced, you can pretty much guarantee that dual single-core dual machines would cease to exist. All the duals would be built around the dual-core chips.

Originally posted by Simon:
Bill Gates ignored that 640k RAM wasn't always going to be enough.

We are in the process of buying dozens of quad Xeon boxes for our lab. Quad cores is not tabu.

Saying we Mac users should not care about it is saying we only need Fisher Price computers.

And how much do those quad Xeon machines cost? No need to look it up. They're often in the range of something like $20000 (if you want the fastest chips, and real support). Of course quad machines are not taboo, but it's quite obvious that Apple doesn't see the need to market these, at least until dual-core chips exist. And no, a dual 2.5 GHz Power Mac is not a Fisher Price computer, nor would be a dual-core 3.0 GHz Power Mac.

BTW, that 640 KB thing is an urban legend:

Question: "I read in a newspaper that in l981 you said '640K of memory should be enough for anybody.' What did you mean when you said this?"

Gates: "I've said some stupid things and some wrong things, but not that. No one involved in computers would ever say that a certain amount of memory is enough for all time."

"Meanwhile, I keep bumping into that silly quotation attributed to me that says 640K of memory is enough. There's never a citation; the quotation just floats like a rumor, repeated again and again."

[UnixMac]

Originally posted by Eug Wanker:
We can pretty much ignore anything about 4-core chip designs, because that ain't gonna happen any time soon in Macs. We can also ignore anything about 4 x dual-core Macs too, since that ain't gonna happen any time soon either.

That may well be true since that's pretty much Sun and SGI's domain (higher end workstations).. but it would be nice to see Apple branch off into that market. They could keep their line of PowerMacs and keep cost at a cap of $3299... and then just like they did the rack-mounted servers... add a line of high end workstations that use the Power5 with larger cache and more than 2 CPU's... prices comparable to Sun and SGI... Tiger would do great with this.

[Eug Wanker]

Originally posted by UnixMac:
That may well be true since that's pretty much Sun and SGI's domain (higher end workstations).. but it would be nice to see Apple branch off into that market. They could keep their line of PowerMacs and keep cost at a cap of $3299... and then just like they did the rack-mounted servers... add a line of high end workstations that use the Power5 with larger cache and more than 2 CPU's... prices comparable to Sun and SGI... Tiger would do great with this.

POWER5 would be extremely unlikely. Dual dual-core POWER5-lite machines (4 cores, but with SMT, seen as 8 CPUs), might happen in a few years, but POWER5 is far too expensive IMO to be a practical consideration.

[Simon]

Originally posted by UnixMac:
That may well be true since that's pretty much Sun and SGI's domain (higher end workstations).. but it would be nice to see Apple branch off into that market.

Of course. The Xserves did extremely well because they offered a lot of Gflops per volume (which is important in server room envirnoments) at an affordable price. Now imagine Apple keeps the current design but uses two 970MPs. That gives you four cores in a 1u unit. You just (at least) doubled your CPU power 'out of the box'. Of course this wouldn't be a mass market product, but it's a product for a very profitable market.

[MindFad]

Drool.

That is all.

[jfaller]

Originally posted by Eug Wanker:
And why would it matter? Because dual-core is a theoretically cheaper way of making duals.

Err... no.

The semi-conductor industry is all about yield. You want to maximized the number of chips that pass at specified current and voltage with the minimum amount of testing and packaging. Dual core chips actually DECREASE your yield. A functioning single core (with N transistors) has a probability p for failure, and a dual core (with roughly 2N) has a probability of ~2p. So you end up throwing away more dice.

I work for a semiconductor company who is using the same IBM fabs that they make G5's, and we can't get IBM to make working 0.9 chips. Their process is all hosed. I wouldn't bet on dual G5's that are affordable to consumers anytime in the near future.

[Eug Wanker]

Originally posted by jfaller:
Err... no.

The semi-conductor industry is all about yield. You want to maximized the number of chips that pass at specified current and voltage with the minimum amount of testing and packaging. Dual core chips actually DECREASE your yield. A functioning single core (with N transistors) has a probability p for failure, and a dual core (with roughly 2N) has a probability of ~2p. So you end up throwing away more dice.

Yeah, except that if one of the two cores works, you can just sell it as a single-core chip. It's not as if Apple will be selling only dual-core based machines. They still have the iMac, single Power Macs, and single Xserves to use those chips.

By the way, wouldn't the probability of failure (of a working dual-core chip) be p�, not 2p?

I work for a semiconductor company who is using the same IBM fabs that they make G5's, and we can't get IBM to make working 0.9 chips. Their process is all hosed. I wouldn't bet on dual G5's that are affordable to consumers anytime in the near future.

Not saying this is true for you, but IBM claims that their successes with chips designed themselves is much better than with 3rd party designs. Mind you, their own 90 nm G5s haven't had the best track record as you said.

[jfaller]

Originally posted by Eug Wanker:
Yeah, except that if one of the two cores works, you can just sell it as a single-core chip. It's not as if Apple will be selling only dual-core based machines. They still have the iMac, single Power Macs, and single Xserves to use those chips.

By the way, wouldn't the probability of failure (of a working dual-core chip) be p�, not 2p?


Not saying this is true for you, but IBM claims that their successes with chips designed themselves is much better than with 3rd party designs. Mind you, their own 90 nm G5s haven't had the best track record as you said.

Actually, about the probabilities...
We need to think of it as Bernoulli trials. I stated the probability of failure as p, so what we're really interested in (when discussing yield) is the probability of passing. so for dual cores, the prob. is:
no failures: (1-p)(1-p) = 1 - 2p + p^2.
As p^2 is typically much less than 2p, we can ignore and say prob. for passing is 1-2p, or the prob. for failure ~2p.

About dual cores:
Typically dual cores aren't laid out well to be recut to a single core die. Most dual cores work with a shared L1/L2/both -- I imagine that you could(?!?), but again you pay on the backend with more test time.

As far as IBM's yield, I think we can all safely look at the G5 and tell that they're lying. In this case, the proof is in the pudding.


----
Edited with prob. adjustment.

[Eug Wanker]

Originally posted by jfaller:
About dual cores:
Typically dual cores aren't laid out well to be recut to a single core die. I imagine that you could(?!?), but again you pay on the backend with more test time.

Not recut. Just blow a few fuses, to completely inactivate the second defective core. ie. The core would still be there in the final chip, just not used.

BTW, I am not an engineer, but this is the little of what I've read on the subject.

[jfaller]

Originally posted by Eug Wanker:
Not recut. Just blow a few fuses, to completely inactivate the second defective core. ie. The core would still be there in the final chip, just not used.

BTW, I am not an engineer, but this is the little of what I've read on the subject.

You probably can use fuses to gate off power/gnd/signal to the dead core, but you've still got all that die there. I imagine (and this is reaching the edge of my training as an engineer -- I don't worry about these issues) that you'd have packaging problems with all that die sitting there. On our chips, packaging is about 50% of material cost.

[Eug Wanker]

no failures: (1-p)(1-p) = 1 - 2p + p^2.
As p^2 is typically much less than 2p, we can ignore and say prob. for passing is 1-2p, or the prob. for failure ~2p.

Oops, I guess what I should have said is the probability of passing is q�, where q = (1-p) = single-core yield. (As you said, the probability of failure being p� is incorrect. )

Anyways, IBM claims they will greatly improve yields this quarter.

[macgfx]

We've already seen Quad CPU PowerMac's. Sure, they came from a "Clone" ,however the same things hold true. Unless MHz stop going up we are unlikely see software that can use these 4 CPU's to full potential make it's way to the consumer level OR by the time we do a single CPU Mac will be faster than the Quad.

I won't discount that Apple would enter into the Quad CPU Market that already exists IF they could offer a lower cost to performance ratio such as the Xserve. I'm not sure how IBM would react to Apple offering a direct competitor to Power-4/5 Blades.

BTW, The PowerMac G5 with it's all to large heatsink/water cooling and heat zone 3 billion fan case with room for only two HD's and one CD/DVD drive looks foolish.

[Simon]

Assuming the 970MP actually materializes and runs at 3GHz, I'm wondering how its heat output will compare to the 970FX. Not just dissipated power in W, but also power flow W/m^2. What kind of cooling requirements will this thing have? Naturally, taking cores apart and putting them on separate circuits gives more space to access for cooling. But on the other hand the design has certainly been improved if they are going to start at 3GHz. What's the tradeoff here?

And secondly, will Apple really use it? After all IBM could be just building it for their own purposes like their 970 blades running Linux. Right from the start, what benefit does the 970MP have over dual 970GXs? Will it be cheaper? In the beginning, probably not. Will it be here earlier? No, the GX is scheduled to arrive first. Will it be cooler? Will it run faster (fast on-chip communication, sharing a larger L2, etc.)? Will the current bus be adequate? Memory access?

[DrBoar]

There is already a IBM 970 with twise the Cache size of the one that Apple use and if that can run at 2.3 GHz in a thin blade it ought to run a bit faster in a G5 tower. So this could come today if Apple shoose to do so, asuming that there are no yield differences in 512 and 1024k 970s.

Getting a integrated memory controller into the IBM970 would benefit the performance in desktop computing (at least the results in DOOM 3 with various AMD CPUs suggest this ) For the embedded market that IBM probably have an interest in IBC is a good thing as well).
IMC should come ASAP. IMC could be inplemented at 130nm and does not require any advances in current technologies as such. However as they probably aim to phase out the 130 nm ones a second major revision at 90nm seem to be a good candiate for IMC.

>3 GHz and dual core and what not will only come after many engineering problems has been solved.

With Jobs promise that the Mac should be at 3 GHz last summer I am sure that Apple would prefer to jump straight to 3 GHz rather than staying at 2.5 GHz with implementations of various performance enhancements. But that really rest with IBM...

I think we should get them 6 more months before raving about Motorola/G4 take two

[Eug Wanker]

Originally posted by Simon:
Assuming the 970MP actually materializes and runs at 3GHz, I'm wondering how its heat output will compare to the 970FX. Not just dissipated power in W, but also power flow W/m^2. What kind of cooling requirements will this thing have? Naturally, taking cores apart and putting them on separate circuits gives more space to access for cooling. But on the other hand the design has certainly been improved if they are going to start at 3GHz. What's the tradeoff here?

Power utilization is the key question, but I really doubt they start at 3 GHz. Maybe their max speed is 3 GHz.

And secondly, will Apple really use it?

Absolutely. No question at all. It's essentially the only option Apple has.

After all IBM could be just building it for their own purposes like their 970 blades running Linux.

Too low volume.

Right from the start, what benefit does the 970MP have over dual 970GXs?

See messages above.

Will it be cheaper? In the beginning, probably not.

Well, don't forget that the motherboards supporting a 970MP could be simpler/cheaper.

Will it be here earlier? No, the GX is scheduled to arrive first.

If the GX is a reject MP, then they could be here pretty much at the same time.

Originally posted by DrBoar:
There is already a IBM 970 with twise the Cache size of the one that Apple use and if that can run at 2.3 GHz in a thin blade it ought to run a bit faster in a G5 tower. So this could come today if Apple shoose to do so, asuming that there are no yield differences in 512 and 1024k 970s.

Where? If you're talking the 2.2 GHz JS20 with 1 MB L2 cache, that's a misprint. They actually have 512 KB L2 cache.

Getting a integrated memory controller into the IBM970 would benefit the performance in desktop computing (at least the results in DOOM 3 with various AMD CPUs suggest this ) For the embedded market that IBM probably have an interest in IBC is a good thing as well).
IMC should come ASAP. IMC could be inplemented at 130nm and does not require any advances in current technologies as such. However as they probably aim to phase out the 130 nm ones a second major revision at 90nm seem to be a good candiate for IMC.

IMC would be great, but so far nobody has reported that it's coming on these 970GX/970MP parts. It's all conjecture of course though.

[Simon]

Originally posted by Eug Wanker:
Power utilization is the key question

It's not just pure power. Apple's current cooling problems arise from the power density (or power flow actually), not the absolute power. Since two cores are now supposed to be on the same die I'd imagine this problem could get even worse. The question is if design changes can overcome this.

but I really doubt they start at 3 GHz. Maybe their max speed is 3 GHz.

I've had the same suspicion myself.

Assuming the GX ships at 2.5GHz, the first question is how much benefit could we expect compared to the 2.5GHz FX. Will it be enough for a full new PowerMac revision? Or could Apple quietly slide it in a running revision like the mixing of FXs and 'original' 970s in the current PowerMacs?

Well, don't forget that the motherboards supporting a 970MP could be simpler/cheaper.

Certainly. But I'm wondering if these savings will make up for the increased cost of the CPU. I could imagine that at least during ramp-up and early production they won't.

If the GX is a reject MP, then they could be here pretty much at the same time.

This is not what the TS article is suggesting. They mention the GX will be ready by first or second quarter 05 and that nothing firm is known about the MP's availability. That doesn't sound like it's the same design and production up to the point where bad MPs are rejected and turn into GXs.

[Eug Wanker]

Originally posted by Simon:
It's not just pure power. Apple's current cooling problems arise from the power density (or power flow actually), not the absolute power.

I agree completely.

Since two cores are now supposed to be on the same die I'd imagine this problem could get even worse. The question is if design changes can overcome this.

Two dies on one core won't really change the power density issues.

This is not what the TS article is suggesting. They mention the GX will be ready by first or second quarter 05 and that nothing firm is known about the MP's availability. That doesn't sound like it's the same design and production up to the point where bad MPs are rejected and turn into GXs.

They could be right, but then what's going to happen to all the reject MPs? I'm just trying to think thru this from a cost perspective. It's pretty clear though that TS doesn't have all the details (although we have even less).

[Simon]

Originally posted by Eug Wanker:
Two dies on one core won't really change the power density issues.

Why not exactly? I guess the packaging (and thus the coolant access) area doesn't necessarily have to double in size just because the cores double.

They could be right, but then what's going to happen to all the reject MPs?

Chips aren't just rejected. They are clocked at different speeds. Rejected pertains to a certain clock speed. The rejected 3GHz MPs could well be the accepted 2.5GHz MPs. The reason you pay more for the higher clocked CPU is this process, but there is in general no design difference between two differently clocked CPUs. As the engineer pointed out in his post above it's not trivial to make single-core ICs out of rejected dual core processors. Even if this might seem like a good idea economically.

[Eug Wanker]

Originally posted by Simon:
Why not exactly? I guess the packaging (and thus the coolant access) area doesn't necessarily have to double in size just because the cores double.

The actual die area of a single dual-core chip is almost double the size of two single-core chips, for obvious reasons. However, power density isn't really an issue for the chip as a whole. The issue specific hot spots, and going dual-core would not change those hot spots (unless they tweaked the design there too). I've read that Intel does a lot more designing by hand, in order to minimize these hot spots.

Chips aren't just rejected. They are clocked at different speeds. Rejected pertains to a certain clock speed. The rejected 3GHz MPs could well be the accepted 2.5GHz MPs. The reason you pay more for the higher clocked CPU is this process, but there is in general no design difference between two differently clocked CPUs. As the engineer pointed out in his post above it's not trivial to make single-core ICs out of rejected dual core processors. Even if this might seem like a good idea economically.

One can think of yield as referring to functional chips, not clock speed. IOW, some chips ARE rejected. ie. They don't work at all, or else don't work at any acceptable clock speed. In a dual-core design, there will be some chips that have one perfectly fine core, and one that would be completely useless at any acceptable speed. Either you throw the entire chip out, or you find some way of using the good core. Making it a single-core chip by blowing some fuses is one solution to this issue. However, you still have to package the defective core, which costs money.

Originally posted by jfaller:
You probably can use fuses to gate off power/gnd/signal to the dead core, but you've still got all that die there. I imagine (and this is reaching the edge of my training as an engineer -- I don't worry about these issues) that you'd have packaging problems with all that die sitting there. On our chips, packaging is about 50% of material cost.

How much do your chips cost though? I'm guessing the chips that Apple gets from IBM (970FX single-core) cost well over $150 a shot. (Don't ask me to back up that price, because I can't. )

[Simon]

Haha, my thread made it to MacNN's front page. Will I get a 970MP as a prize now? If not, of course I'd also settle for a double quarter pounder w/ cheese and bacon.

[olePigeon]

How about a Quad Processor G5? 4x 3GHz. Mmmmmm.

[Simon]

Originally posted by olePigeon:
How about a Quad Processor G5? 4x 3GHz. Mmmmmm.

Sounds good to me. That's what the Xserve should offer. I'm just wondering what kind of bus and system controller you'd need in order to really be able to make use of those 4 cores.

[Eug Wanker]

Originally posted by Simon:
Sounds good to me. That's what the Xserve should offer. I'm just wondering what kind of bus and system controller you'd need in order to really be able to make use of those 4 cores.

Why should an Xserve, a low end 1U server, have 4 CPUs? I'm not sure there's really much call for it at this time.

[Simon]

I just knew somebody would come here and try to explain why Apple doesn't need to capitalize on the G5's incredible power. Duh.

Originally posted by Eug Wanker:
Why should an Xserve, a low end 1U server,

It shouldn't remain either low-end or 1u. At least not exclusively.

I'm not sure there's really much call for it at this time.

As long as people buy such systems from other companies running other OS' there is. And there are plenty of people doing that. Apple wants to be in science. They advertise it all the time. Well here science is, telling them what they need to offer. And at the same time, there is hardly another market segment where Apple could have similarly high profit margins.

The current Xserves are nice, but actually they're hardly more than a PowerMac G5 in a 19" rack-mountable enclosure. They are fine. But they need some bigger brothers. More cores, more memory. Basically, the bus should be the limit. That said, they need a better bus too.

[Eug Wanker]

Personally, I think Apple should start slow and build up their server sales before they try to tackle the mid/high end of server sales. The problem is that there is little point in selling hard core servers if there isn't the support available. Unfortunately, Apple does not have the same type of support infrastructure in place that other companies do. The reason people buy other-than-Apple, is because of the support and such, not simply because of CPU performance. And indeed, it seems that slow and steady is exactly Apple's approach.

Indeed, I'd guess there'd be a much larger quad Power Mac market than any market for quad Xserves. Start there first.

[Simon]

In science, especially in large research facilities, we support ourselves. We don't use our severs like a bank does. All we need are hardware parts (and software documentation). Apple has plenty of that.

It's the hardware and its performance per m^3 that counts. Apple could have its foot in the door easily here.

That said, I agree they should push the PowerMac at the same time. Absolutely.

[Eug Wanker]

Originally posted by Simon:
In science, especially in large research facilities, we support ourselves. We don't use our severs like a bank does. All we need are hardware parts (and software documentation). Apple has plenty of that.

It's the hardware and its performance per m^3 that counts. Apple could have its foot in the door easily here.

For the scientific market, Apple has chosen instead to go the small compute cluster approach. I agree with this approach (for now). It serves part of the market looking for self-supported hardware in the mid-priced range, without Apple having to create orphaned $30000 all-in-one machines. Apple's value added component is the pre-installed software.

[OreoCookie]

I'm not sure about 3 GHz. The dual-core Opterons will probably feater a lower clock to be in the same power envelope, i. e. lower clock speed to be able to run at lower voltage so that you can have two cores dissipate as much heat as one core running at higher GHz and voltage.

[UnixMac]

I for one would REALLY like to see Apple continue to grow as a player in the high-end computer world. It's lends credibility to OS X as a premier Unix platform, and it also has a "flagship" prestige factor that can bleed down to their lowest eMac.

Just think if they sold 10-15% of all of the clusters sold in the market... We can finally then stop talking about Apple as a 3% of the market computer... OS X is key to this, and the continued development of X, in the 64 direction, with the right amount of capital investment on Apple's behalf will unseat Solaris and Linux if they play their marketing cards right.

[Eug Wanker]

Originally posted by OreoCookie:
I'm not sure about 3 GHz. The dual-core Opterons will probably feater a lower clock to be in the same power envelope, i. e. lower clock speed to be able to run at lower voltage so that you can have two cores dissipate as much heat as one core running at higher GHz and voltage.

I'm optimistic for a 3 GHz 970GX. A 3 GHz 970MP would also be possible, although it would be very high power chip and it would be difficult to get good numbers of dual-core chips at 3 GHz.

Originally posted by UnixMac:
Just think if they sold 10-15% of all of the desktops sold in the market... We can finally then stop talking about Apple as a 3% of the market computer...

Fixed.

[UnixMac]

Originally posted by Eug Wanker:
I'm optimistic for a 3 GHz 970GX. A 3 GHz 970MP would also be possible, although it would be very high power chip and it would be difficult to get good numbers of dual-core chips at 3 GHz.


Fixed.

Are you basically disagreeing with my theory that the key to more desktops is sales of more enterprise systems? Or did you misread my intent?