Dual Core PowerPC 970s

[MikeBike]

The result is that we will see real world performance gains far exceed the simple scaling of CPU number and speed. I think the next generation of G5's will probably use a single CPU that is dual core, making the machines more affordable.

The future looks bright for the Mac, if you ask me.

HAY! I don't want to hear ANY talk about dropping 2 chips for 1, except in a low end model, in the Power Mac line. I want, no demand, to see, Dual-Core, Dual-Processor machines! I want my QUAD processor. Don't you wreck my dreams man.

 

[maxvamp]

Bring my QUAD iMac!!!

Actually, I am fairly positive that iMacs, iBooks, and eMacs will not see a dual core. I would be surprised if the iBook even got a G5 any time soon, leaving the PB space only at a single core G5.

I mentioned it before, but here comes the bummer... Unless AMD starts pushing dual dual cores, I would more likely expect the PM to drop to one dual core chip, and the dual chips stay with XServes where the Scientist will make the best use of them.

I WANT MY XServe ???

Max.

 

[Phinius]

What's your source? The articles that I have read seem to indicate the multi-core 970 will be introduced around 3GHz, not top out there.

The 970 is based on the Power4 core. There is very little possibility that the PowerPC that is based on the Power core will pass the Power chips in frequency on the same process size. The Power5+ is expected to peak at 3GHz on a 90-nm process. Therefore, the 970 multi-core will not go beyond 3GHz on a 90-nm process size.

I think the 970mp is designed to replace the need for 2 individual CPU's, so if the next revision is 3GHz, then the multi-core 970 will be 3GHz. Now, I could be wrong. I'll admit that. But I don't think any of the above is far off the beaten path and is actually likely.

The reason for a dual-core is not only to move two separate processors onto one chip, but also to get more performance out of every chip made. Apple could use dual-channel memory for two chips as they do now and have two additional processors too boot.

I don't care how much spin you put on it, unless they stop indicating GHz ratings on their computers there is no way Apple will release a machine with lower GHz or equal GHz as they have now and call it an upgrade-- no matter what the real world performance is. It's a markting nightmare.

A IBM executive recently stated that scaling was dead and that increases in performance will have to come mainly from something other than frequency increases. Going to a dual-core chip, and adding SMT (Power5) are just two of the ways that IBM is adding performance without increasing frequency.

 

[Phinius]

Actually the PPC 970 has always had 512KB of L2 cache. The 970MP would double this to 1MB per core.

Your right. I miscued there.

2.5Ghz to 3Ghz is a %20 increase. We will not be able to accurately compare the 2.5Ghz Duals to a 3Ghz Dual Core because of the cache size differences and faster chip to chip links in a dual core system. Also it seems that the 970MP may have lengthened some pipelines which would allow it to clock higher but would change the IPC.

Don't expect the 970MP to go beyond 3GHz. It's unlikely that IBM will extend the 970 pipeline stages beyond what exists now. That's due to the 970 core being based on the Power4. Even the Power5 core does not have more pipeline stages than the Power4.

 

[ddtlm]

MikeBike:

Why would a Power5-lite be smaller than a 970( based upon the Power4 ? ).

One core vs two.

More cores also means more cache: each cpu will have it's own independent cache, and a properly written os should be able to keep a much higher level of cache "coheriency"?

Hard to say for sure, but I bet the trouble with keeping all the caches coherent overrules the benefit from there being more total cache space.

Phinius:

The upcoming 90nm G4 won't have a 166Mhz FSB. I have seen a internal Motorola document that describes the next G4 as having DDR and DDR2 capability.

And I linked to their web page where they declare in no uncertain terms that the e600 is pin-compatible. That means the same FSB. (Besides, one can claim current G4's support DDR memory... through the system controller.)

The 970 is based on the Power4 core. There is very little possibility that the PowerPC that is based on the Power core will pass the Power chips in frequency on the same process size. The Power5+ is expected to peak at 3GHz on a 90-nm process. Therefore, the 970 multi-core will not go beyond 3GHz on a 90-nm process size.

I don't see why a Power5 core can be used to judge the Power4-based 970, and in any case, the 130nm 970 did clock higher than the Power4 which topped out at 1.9ghz (130nm) so far. But yeah I'd agree with what your saying about the 970mp not making it past ~3ghz.

 

[Phinius]

And I linked to their web page where they declare in no uncertain terms that the e600 is pin-compatible. That means the same FSB. (Besides, one can claim current G4's support DDR memory... through the system controller.)

The G4 is moving to 2GHz and if you believe that it will still have a 166MHz bus, then you are in for a surprise. The G4 has a history of moving to a faster bus speed with the same pin count.

Also, the G4 currently only uses half the speed of DDR memory, even though Apple chooses to use for marketing reasons.

I don't see why a Power5 core can be used to judge the Power4-based 970, and in any case, the 130nm 970 did clock higher than the Power4 which topped out at 1.9ghz (130nm) so far. But yeah I'd agree with what your saying about the 970mp not making it past ~3ghz.

The Power5+ is expected to reach 3GHz on a 90-nm process and Steve Jobs announced that the 970FX would reach 3GHz. Doesn't that alone indicate something? I believe it's more than coincidence and no the Power5 does not have more pipeline stages than the Power4, so it is relevent to compare the topend frequency of the Power5 to the 970FX on the same process size.

Incidently, there should be about a 10-15% performance increase from doubling the 970 L2 cache and up to a 50% increase in performance from adding another processor to the chip. That on top of the 50% increase in frequency going from 2GHz to 3GHz.

The next round of improvements should either include a move to a Power5 based PowerPC chip on a 90-nm process or 65-nm process. That might give a performance boost of up to 30% for the SMT and another 30% from the process shrink.

It looks like Apple and IBM are aiming for about a 50-60% speed improvement for each version of the 9XX PowerPC chip that Apple is using.

 

[Frobozz]

The 970 is based on the Power4 core. There is very little possibility that the PowerPC that is based on the Power core will pass the Power chips in frequency on the same process size. The Power5+ is expected to peak at 3GHz on a 90-nm process. Therefore, the 970 multi-core will not go beyond 3GHz on a 90-nm process size.

Bah! The only scenario the above makes sense is a dual cpu dual-core configuration. I don't think we'll see it. If we do, it'd be held for all but the xServe and perhaps a high end PM G5 config.

I see your point, I just think the single CPU / dual-core at 3GHz as being far more likely. Aside from the cost reduction (?) the marketing required to remove GHz from the equation would be interesting. Think of it this way: if Apple releases a CPU, no matter how fast the cache, no matter how many cores, that is less than 3.0 GHz they will fail in the public's eye. I agree that it amounts to basically nothing in real world performance, as a 2.5 GHz chip or 2.8 GHz chip can easily outperform a faster clocked chip-- but Apple will have to make a big shift in it's markting like AMD and Intel are starting to do.

I agree that scaling is no longer the most viable speed increase for a CPU, too.

 

[Fukui]

The 970 is based on the Power4 core. There is very little possibility that the PowerPC that is based on the Power core will pass the Power chips in frequency on the same process size. The Power5+ is expected to peak at 3GHz on a 90-nm process. Therefore, the 970 multi-core will not go beyond 3GHz on a 90-nm process size.

Power4/5 are designed with reliability in mind more than just raw clockspeed. Example being thier thicker oxide gates which reduces clock speed but it can take higher temperatures, the 970 is more for higher clock speed but it may die earlier than a power4/5 based system, would you still be using your G5 after 15/20 years? 30? If the 970 passes the power4/5 in clock speed, thats fine, people who buy Power 4/5s need them for a reason, same with the G5; its different markets. We need speed, they need reliability+speed but reliability wins.

 

[wdlove]

I enjoy reading this thread, but it is well above my head. Don't understand all the abbreviations. I do agree that reliability is very important factor.

 

[ddtlm]

Phinius:

The G4 is moving to 2GHz and if you believe that it will still have a 166MHz bus, then you are in for a surprise. The G4 has a history of moving to a faster bus speed with the same pin count.

Hah, so 200mhz? 233mhz? It's not gona scale much. Pin compatible.

The Power5+ is expected to reach 3GHz on a 90-nm process and Steve Jobs announced that the 970FX would reach 3GHz. Doesn't that alone indicate something?

No.

I believe it's more than coincidence and no the Power5 does not have more pipeline stages than the Power4, so it is relevent to compare the topend frequency of the Power5 to the 970FX on the same process size.

Ah, now I haven't gone counting pipeline stages but yeah those could mean something... though we are talking about different cores so comparison of stages alone really doesn't settle it.

 

[FFTT]

Thinking to myself... Ignorance Is Bliss


This is enough to make anyone crazy.

2 weeks ago, the 2.5 would have been my dream machine.

Now that it's all too clear the individual dual processors
may soon go the way of the Dodo, I'm reluctant to buy anything.

Water cooled DOES seem to be much safer than trusting 9 cooling fans
for long term survival. Otherwise, I would have gone for a 2 GHz dually
months ago. I still can't help but worry about the long term effects
of so much heat causing premature system failures.

I just hope now that a dual core dually WILL be available soon.
And that long term reliability will not be an issue.

I am hopeful that the upcoming G5 iMac will give us a few more clues as to where things are going and how soon.

 

[maxvamp]

FFTT

Thank you for the perspective.

I would love to have a Dual 2.0, but realistically, the next time I buy a Mac, I will be buying the fastest possible.

Why you ask?

Well, I have had un unfortunate luck of Macs lasting to damn long. I own 5 machines, which I replace / upgrade as they wear out or become unusable. As you can see below there seems to be too much longevity with Macs:

Main Servers

PM G3 300 ( CVS Store and File / Print server ) running Panther
Dual 1GHz PIII - Backup server and dev box ( I know, bad combo )

Workstation - Games

PM 733 DA - General stuff / video / dev / other ( GoBan )
Home brew Water Cooled AMD64 3000+ XP running games ( For the kids )
iBook G3 600 test machine / portable machine / internet box / Games

Due to the fact the Macs are still very productive ( damn Macs ) My next machine may end up being a Dual-Dual PM or G5 PB 17"

When it comes to faster ( Mac ) machines, what do you really need?

Max.

 

[MikeBike]

Power4/5 are designed with reliability in mind more than just raw clockspeed. Example being thier thicker oxide gates which reduces clock speed but it can take higher temperatures, the 970 is more for higher clock speed but it may die earlier than a power4/5 based system, would you still be using your G5 after 15/20 years? 30? If the 970 passes the power4/5 in clock speed, thats fine, people who buy Power 4/5s need them for a reason, same with the G5; its different markets. We need speed, they need reliability+speed but reliability wins.

I want to emphasize fukui's point.
An Enterprise Server is made immune to Cosmic Ray abuse.
Wider Traces, ECC Parity Checked memory so that your long running server isn't getting it's OS clobbered in memory while it's running. And your financial apps are accurate, and you aren't wasting your debug time on sporadic errors. These are the POWER chips.

Your consumer chips ( 970, FX ) don't have the wide traces, and ECC memory. As a consequence, they can clock HIGHER. But, you are not protected from sporadic errors, but, you wouldn't PAY for that kind of protection anyway. So, if the Power5 will clock to 3Ghz, then the 975? should go to at least 3.5Ghz.

 

[Phinius]

Your consumer chips ( 970, FX ) don't have the wide traces, and ECC memory. As a consequence, they can clock HIGHER. But, you are not protected from sporadic errors, but, you wouldn't PAY for that kind of protection anyway. So, if the Power5 will clock to 3Ghz, then the 975? should go to at least 3.5Ghz.

Ahh, but the 970 chips do have the same thickness of traces that the Power chips do. IBM didn't change it when they manufactured the 970.

The 970 does use ECC memory.

 

[Phinius]

Hah, so 200mhz? 233mhz? It's not gona scale much. Pin compatible.

Internal Motorola documents state that the 2GHz G4 will have DDR and DDR-2 capability, so you can double your bus frequencies above.

Ah, now I haven't gone counting pipeline stages but yeah those could mean something... though we are talking about different cores so comparison of stages alone really doesn't settle it.

The Power5 is followup to the Power4 and it doesn't have more pipeline stages than the Power4. Which means that IBM is getting the performance improvements from something other than frequency boosts coming from architectural changes. The Power5 will move to 3GHz from a process shrink and the use of strained silicon.

Apple does not have the sales volume, nor does IBM have the money losing gift giving heart to create a special higher frequency architectural core just for Apple's small marketshare. The 9XX PowerPC chips will have the same amount of pipeline stages as their Power4 or Power5 brothern. There will be a substantial frequency boost when IBM comes out with the Power6 in 2006, however. It's expected to reach 5GHz. Which would probably put the pipeline stages at about the same amount as the original Pentium 4.

 

[ffakr]

Ahh, but the 970 chips do have the same thickness of traces that the Power chips do. IBM didn't change it when they manufactured the 970.

i'm fairly certain that this is not the case. Do you have any evidence that the Power 4 and the 970 use traces of the same thickness? I'll look for something specific to contradict you later.. I'm a bit busy now.

The 970 does use ECC memory.

The 970 doesn't care what memory you use, the memory controller isn't on the chip. There are, however, NO macs with 970 chips that support ECC. The 970fx chip in the xServe supports ECC, but that is because the system chipset supports ECC memory.
I think the original poster of the ECC comment was making a broad point about the quality of enterprise servers vs. a Mac desktop.

 

[Phinius]

i'm fairly certain that this is not the case. Do you have any evidence that the Power 4 and the 970 use traces of the same thickness? I'll look for something specific to contradict you later.. I'm a bit busy now.

If I remember correctly, my traces thickness information comes from a series of 970 articles at Arstechnica.com where the author interviews a few IBM executives.

 

[MikeBike]

Ahh, but the 970 chips do have the same thickness of traces that the Power chips do. IBM didn't change it when they manufactured the 970.

Prove it.

It's time you did some research.
You obviously haven't been following the IBM/Apple press releases since the first G5 was announced and put into a Mac product. Why don't you look up how Server chip makers like Sun, HP and IBM build their products to protect against Cosmic Ray error hits, and then see if those characteristics are in Any consumer product from Intel, AMD or Apple.

 

[FFTT]

Speaking of research...

/http://www.microsoft.com/presspass/features/2002/jul02/0724palladiumwp.asp

Reading this made me start to wonder what Apple might have up their
trees and how the newer system architectures may further limit the users
freedom by installing DRM on a chipset level.

Would this be avoidable by purchasing the current systems?

Just asking

 

[ddtlm]

Phinius:

Internal Motorola documents state that the 2GHz G4 will have DDR and DDR-2 capability, so you can double your bus frequencies above.

Hmmm, the words "pin compatible" sound familiar? Also what's your fascination with repeating things to me? I heard your claims about DDR the first time, and I told you at even current G4's support it... through the chipset. They'll support anything through the chipset, in fact.

The Power5 is followup to the Power4 and it doesn't have more pipeline stages than the Power4. Which means that IBM is getting the performance improvements from something other than frequency boosts coming from architectural changes.

Those kinds of changes can effect clockspeed still, either up or down. You cannot count pipeline stages and draw solid conclusions.

Apple does not have the sales volume, nor does IBM have the money losing gift giving heart to create a special higher frequency architectural core just for Apple's small marketshare. The 9XX PowerPC chips will have the same amount of pipeline stages as their Power4 or Power5 brothern.

This doesn't address anything I said, nor does it address anything I thought to be true. I think the root cause of your confusion is facination with the number of pipeline stages. A lot of variables effect clockspeed, number of stages is only one of them.

 

[FFTT]

Prove it.

It's time you did some research.
You obviously haven't been following the IBM/Apple press releases since the first G5 was announced and put into a Mac product. Why don't you look up how Server chip makers like Sun, HP and IBM build their products to protect against Cosmic Ray error hits, and then see if those characteristics are in Any consumer product from Intel, AMD or Apple.

Tech update:

Tin foil hats will be provided with all new G5 iMacs

 

[ffakr]

If I remember correctly, my traces thickness information comes from a series of 970 articles at Arstechnica.com where the author interviews a few IBM executives.

here is the link to the interview with Peter Sandon..

http://arstechnica.com/cpu/03q2/ppc970-interview/ppc970-interview-1.html

I don't see anything related to the thickness of the traces.

From everything I've seen, the Power5s continue to be manufactured in a more robust manner than the 970s.

Do you have any proof of your contention that the Power 5 is not designed to be more robust than the 970?

 

[wdlove]

Is it becoming the opinion of our technically knowledgeable members that the 2.5 should not be purchased. So would it be wise to wait for this new dual core? Do you expect them to be announced during MWSF '05?

 

[maxvamp]

wdlove - If you wait...

You'll get none!!!

There always has been and always will be speculation on the next great thing.

People often wait for the next great thing, and never buy anything because as soon as they are ready to buy, something better is speculater to be just around the corner.

If you are in the market for the biggest and baddest Mac around, buy a 2.5 now. There is nothing wrong with it. If you will be ready for a new Mac next year, buy one then.

Don't base your purchase on what may be coming...

Max.

 

[wdlove]

You'll get none!!!

There always has been and always will be speculation on the next great thing.

People often wait for the next great thing, and never buy anything because as soon as they are ready to buy, something better is speculater to be just around the corner.

If you are in the market for the biggest and baddest Mac around, buy a 2.5 now. There is nothing wrong with it. If you will be ready for a new Mac next year, buy one then.

Don't base your purchase on what may be coming...

Max.

Thank you Max, so your not anticipating anything better at MWSF?

My current Power Mac G4 Dual 450 will be 4 years old September 1st.

My wife has been encouraging me to purchase, she is afraid that my Mac will quit soon. I'm ambivalent, when I hear about the dual core.