Dual Core PowerPC 970s

[zac4mac]

Thanks guys -

Aiden, Sun Baked and Morpheus(by proxy) et al, for a very interesting read.
Will we ever see this thing in a Mac - who knows??

I got quite a performance leap going from a DP550 G4 to a new G5 - 2x2 last year and I don't see a need, personally, for more power soon. But "soon" is a relative term.
The difference, on current equipment, between single and dual procs is significant. If indeed, more is better, then let's have it.

Z

 

[macsrus]

macsrus:
Actually NT supported 4 processors way back in Windows NT 3.1 Advanced Server.....
And Believe it or not Unisys was building 32 processor NT servers using Intel pentiums back in 1995/96 (during the NT 3.51 days)

me:
Yes at 3 they WERE supporting four. But when it first came out, you know, ver 1.0, they were supporting two.

NT 3.1 was the first Windows NT (There wasnt a verson 1.0 thru 3.0)


Another thing Microsoft NOT IBM wrote OS2.... IBM was tricked by Bill and company again on this one..... They wrote OS2 while writting NT at the same time as a replacement.

IBM got mad and took all of the OS2 source code and cancelled the contract with MS for OS2 development....

 

[relimw]

The next machines will have PCI Express (PCIe) graphics, no doubt.

Ahh, now see I'm not too much on my PCI standards these days, how much faster is PCI Express over AGP8x and the various other PCI incarnations?

 

[relimw]

Well, if you look at military classifications, X usually means "experimental". I realize that most people probably would overlook that, but I think the G is cool. I don't know what it actually means, but I kind of think of it as "generation", since one could make the argument that the PPC 601/603/603E was a kind of G1, and the 604/604e was a sort-of G2.

I know, they should just start naming it: IS1, IRS2, IRRS3, etc (Intel S*cks, Intel Really S*cks, etc) ;p

 

[relimw]

more power! Great! Wait! so few really have that serious a need for that kind of power, software just isn't keeping up. More than that, c'mon! Apple's alreaedy got the graphics crowd, etc.

But yet some people *really* need that kind of power. Like those that are building supercomputer clusters, or those that are clustering for mainframe-like, multiple user situations. By now you should know that more than just yourself can make use of your Mac computer if you want/need too. I'd love to have such a beast to work, it'd make things a lot faster, and simpler for myself, and my multiple users to work

Yet those that 'just' check email, or do simple word processing, can go back to using the old DOS-based machines or equilivant, since really, you don't need that much power if the software you are using is coded efficiently. I mean, really, who the heck needs all the bells and whistles that openoffice or M$ office has? Down with the stupid talking paperclip! Long live DOS!

 

[Trekkie]

Excellent, this is what they should be doing. The 1MB caches would be wonderful, 512k is yesterday's high-end, not today's.

Not necessarily.

Cache is used to speed up the processor by keeping it in cache and not in main memory because main memory was slower than cache by orders of magnitude.

However if you have kick ass speed to main memory (like, oh, 1.25GHz the PM G5 2.5GHz does) cache gets in the way in certain workloads, so keeping the cache smaller is not a bad thing....

 

[relimw]

I really don't like this thinksecret article. Sure having a dual Core PowerPC chip would be great but what it all boils down to is unsubstantiated rumors. Anyone could say that IBM is working on Dual Core chips and be right as they probably are, as every other company is doing the same.

Well, first off, IBM is working with, and producing dual core chips, just never before in the lowly PowerPC line. We'll just have to wait and see when they do come out. I'd like a dual core, dual proccessor 3GHz system announced in March '05, shipping immediately Will I get it from Apple? Prolly not

 

[MarcelV]

Tiger Compatibility Support?

...now if only OS X would take advantage of all that 64 bit goodness I'd be giddy as a school girl.

They are able to start production in January. Consider the timing of Tiger, which will have full 64-bit support. I think they will build in support for Dual Core. I think 2004 and 2005 will be very interesting years for Apple's hardware development.

 

[AidenShaw]

Up to 4x faster

Ahh, now see I'm not too much on my PCI standards these days, how much faster is PCI Express over AGP8x and the various other PCI incarnations?

AGP8X peaks at 2.1 GB/sec.

PCI Express (PCIe) is like HyperTransport - different width busses can be used for differnt needs.

The current PCIe chipsets use "16x" width for PCIe graphics, giving a peak of about 8 GB/sec. (4GB/sec up, 4GB/sec down) 32x is also in the spec, for 16 GB/sec.

So it's double to 4 times AGP8X, depending on whether your load is bidirectional.

Other PCI speeds:

o 32-bit/33 MHz - 133 MB/sec
o 32-bit/66 MHz - 266 MB/sec
o 64-bit/33 MHz - 266 MB/sec
o 64-bit/66 MHz - 533 MB/sec
o 64-bit/100 MHz - 800 MB/sec (PCI-X)
o 64-bit/133 MHz - 1066 MB/sec (PCI-X)
o 64-bit/266 MHz - 2133 MB/sec (PCI-X 2)
o 64-bit/533 MHz - 4266 MB/sec (PCI-X 2)

PCI Express:

o x1 - 250 MB/sec bidirectional - 500 MB/sec total
o x2 - 500 MB/sec bidirectional - 1000 MB/sec total
o x4 - 1000 MB/sec bidirectional - 2000 MB/sec total
o x8 - 2000 MB/sec bidirectional - 4000 MB/sec total
o x16 - 4000 MB/sec bidirectional - 8000 MB/sec total
o x32 - 8000 MB/sec bidirectional - 16000 MB/sec total


Unlike HyperTransport, though, PCIe is also a card interconnect - you'll be able to get PCIe network, I/O and other cards. Initially high end cards will have this (you can find InfiniBand, 10 Gbps Ethernet, FibreChannel PCIe cards today) and a PCIe <-> PCI bridge will be used to put both PCI and PCIe slots on the motherboards.

 

[relimw]

AGP8X peaks at 2.1 GB/sec.

PCI Express (PCIe) is like HyperTransport - different width busses can be used for differnt needs.

The current PCIe chipsets use "16x" width for PCIe graphics, giving a peak of about 8 GB/sec. (4GB/sec up, 4GB/sec down) 32x is also in the spec, for 16 GB/sec.

So it's double to 4 times AGP8X, depending on whether your load is bidirectional.

Other PCI speeds:

o 64-bit/100 MHz - 800 MB/sec (PCI-X)
o 64-bit/133 MHz - 1066 MB/sec (PCI-X)
o 64-bit/266 MHz - 2133 MB/sec (PCI-X 2)
o 64-bit/533 MHz - 4266 MB/sec (PCI-X 2)

Yikes! That is a heck of a lot faster than what we're use to on the Mac. So I'd guess that skipping PCI-X 2 would be in Apple's (and our's) best interest. I wonder tho, how many changes would Apple be willing to commit too at one time? Maybe at the next refresh, up the machines to PCI Express and a small speed bump? Then the follow-on refresh would be dual core processor(s) and some other wizzbang snazzy feature?

 

[G4-power]

apparently you are not familiar with IBM's PowerTune technology. They don't need to design a mobile processor if they can get it working correctly, as it will effectively be better than a seperate mobile design, as it will be scalable from 1/64 of clock frequency to full frequency in 3 cycles. this will cut heat and power consumption drastically, because the processor will only be running hard when you're doing heavy computing, and will enter a deep nap mode in between., i.e. when you're typing etc.

Heh, wonder how PowerTune scales when doing massive computing like Folding@Home... Will it use 100% processor speed (and burn your thighs) with a process that uses only free CPU cycles? Or 50%? The maximum heat should still be quite low, that you won't burn your thighs with F@H.

 

[Reality]

I need to agree

It says both cores will share a 1ghz bus...right now each CPU in a dual 2.5 gets 1.25ghz of it's own bandwith, but with that each 3ghz core would have to share only 1ghz, isn't that a step backwards? Or are we getting dual dual-cores?

That point discredits this story. They would not take a step back like that, I hope not anyways.

No dual cores anytime soon guys, sorry.


Visit my site, http://reality.spymac.net
Red Pipe Backgrounds, spice up your desktop today.
Free 30 inch desktops.

 

[gekko513]

AGP8X peaks at 2.1 GB/sec.
... PCIe graphics, giving a peak of about 8 GB/sec...

However, PCIe has higher latency than AGP. That is the reason why current graphics card perform equal or worse on PCIe compared to AGP.

 

[Masao[RY]]

Maybe they'll call it the G5.2.

It would make sense since it is still a 970, dual core, and the third version of the chip (the original and then the 90nm version).

G5se, G5fx, G5 Pro, and so on. Apple has a lot of room to play with the marketing a bit. I do hope they don't call it a G6 since it is way to soon for such. I'd prefer G5fx or something of that nature. Anyway, if this bares fruit I will just have to work like a mofo one summer just to get one

 

[gekko513]

That point discredits this story. They would not take a step back like that, I hope not anyways.

No dual cores anytime soon guys, sorry.


Visit my site, http://reality.spymac.net
Red Pipe Backgrounds, spice up your desktop today.
Free 30 inch desktops.

It's not really a step backward. Dual channel DDR 400 isn't able to fill the 1GHz bus. I think the only data-traffic that's able to fill the current bus capacity is cache lookup on the second processor on a dual setup. The current config has 512KB cache on the processor + 512KB on the other processor.

If they set up that Dual dual-core 3GHz system with 1GHz bus, then each processor will have 1MB of cache by itself + 1MB from the second core on the chip.

So the current config has 512KB of high-speed cache + 512KB@1.25GHz, but the described future system will have 2MB of high-speed cache + 2MB@1GHz. So it would be a huge step forward.

I'm assuming now that one core is allowed to peak at the other core's cache, at least for reads, I think that's the way it works in todays dual configs.

 

[G4-power]

G5se, G5fx, G5 Pro, and so on. Apple has a lot of room to play with the marketing a bit. I do hope they don't call it a G6 since it is way to soon for such. I'd prefer G5fx or something of that nature. Anyway, if this bares fruit I will just have to work like a mofo one summer just to get one

Yeah, we'll see. But I don't think it looks cool, "PowerMac G5fx". The "fx" is just not clear enough. And since there is already a fx in use, the PPC 970fx, that would confuse things a bit.

 

[maxvamp]

Melgross, Aiden, and MacsRus

Melgross and Aiden,

You both are trying hard to take the original comments and skew them out of context to make an untrue point.

Threaded tasking happens all the time in many programs. Mail.app collects mail from multiple mailboxes at once while indexing and doing other tasks. Safari does not download graphics sequentially. Framworks used to build a program, such as a UI **is** part of that program.

DOS is dead. People expect more from programs.

We must agree to disagree at this point, as I doubt that we will change each other's minds.

MacsRus :

One day, we'lll have to start an OS/2 thread, as I am sure that the messy history of this doomed, yet excellent OS would take 200 posts by itself.

Thanx all for the lively debate. I look forward to doing it again.

Max.

 

[itsa]

Now if we could just get them to offer G5's with a decent graphics card.
The least they can do is sell them without cards so you can use that money for one worth having. It sucks that they force you to buy one that you don't even want.

 

[AidenShaw]

Threaded tasking happens all the time in many programs. Mail.app collects mail from multiple mailboxes at once while indexing and doing other tasks. Safari does not download graphics sequentially. Framworks used to build a program, such as a UI **is** part of that program.

I don't think that we disagree that much. I specifically mentioned multi-threaded (overlapped) I/O (mail/safari) in my posts. I also cited video as an example of an application that lends itself to threading across multiple CPUs.

In none of your examples, however, is there any hint at exploiting multiple CPUs.

Does mail.app collect mail faster on a dual? Not appreciably, the task is network and I/O limited!

Does Safari download graphics faster on a dual? Again not significantly - it's a network limited problem.

The original post was talking about the often hard problem of taking a compute intensive task and breaking it into multiple threads to run simultaneously on multiple processors, thereby completing the overall job much faster.

Your statements about overlapped I/O and threaded UI frameworks are true, but you're also skewing the original point with unrelated examples.

 

[ClimbingTheLog]

X.4 == Tiger

(Or is OSX.4 your shorthand for Tiger?

That seems pretty easy to dissect:

Mac OS X 10.4
(we know it's mac here)-->
OS X 10.4
(factor out redundant 10)-->
OS X.4
(take out space)-->
OSX.4

A nice shorthand - of course Tiger is the same number of characters...

 

[ClimbingTheLog]

You mean like "Windows XP Pro"? Yah, I agree, 'X' is used way too much nowadays - OS X is quite enough for me, thank you very much...

No, XP is just egotistical arrogance from Microsoft.

XP is the traditional Christian abbreviation for "Jesus Christ", from Latin.

Microsoft constantly thinks of itself as the underdog; see Cringely's writings on the subject.

So, Windows XP, the first decent core OS aimed at consumers was going to be Microsoft's Saviour, or "the second coming" of Windows.

So, Microsoft Windows XP.

'P' clearly doesn't stand for 'Professional', since we have 'Microsoft Windows XP Professional' as a marketed product.

 

[ClimbingTheLog]

OS X supports two logical processors. It's possible, from what I've read, that 10.4 will support four. The idea that it's easy to continue support up to sixteen, or higher, is wrong.

It's not easy but Mach is purpose-built for it. Mach is the layer under Unix on Mac OS X that does the interfacing with hardware, manages processes, memory, etc. It's been used for Massively Parallel computing, for instance it ran the Cenju-4 MPP machine from NEC which had 1024 processors.

That's not to say you can just load 1024 processors into a Mac - there are bsd kernels, compile-time options, gcc work to be done, etc, but the foundation is there to be built upon.

If Tiger does ship with the reentrant kernel things will get much more interesting.

 

[ClimbingTheLog]

As of OSX.3, the OS supports at least two physical processors. I am not aware that it supports any logical ( hyper threaded ) processors.

If SMT works on the Mac like it does on Intel the OS won't know there's an SMT processor - the BIOS presents it as two processors.

 

[ClimbingTheLog]

Is mail.app collect mail faster on a dual? Not appreciably, the task is network and I/O limited!

On my 1.2 GHz G4 AppleMail is CPU bound. It downloads much faster than it indexes. If I had 4 or 8 CPU's, real or virtual it would busy one with fetching, leave UI on another, and spend the rest indexing, ideally.

It looks like Mail in Tiger will use OS-level indexing which should also help with speed, assuming it's based on V-Twin.

 

[AidenShaw]

If SMT works on the Mac like it does on Intel the OS won't know there's an SMT processor - the BIOS presents it as two processors.

The BIOS also creates a SRAT (Static Resource Allocation Table) as part of the ACPI enumeration. The SRAT contains information that describes physical and logical processors.

If the system doesn't look at the SRAT, you are right that it appears to be two physical processors. (That's why Windows 2000 sees and runs with the two processors - it doesn't know about the SRAT.)

The application programmer can check the GetLogicalProcessorInformation API to learn which processors are real, which are logicals sharing the same real, where memory is connected (in a ccNUMA system),....