PowerPC 970 Redux: Dialogue and Addendum

[MacRumors]

Arstechnica posted more details on the PowerPC 970 after having had the opportunity to chat with two IBM engineers behind the PowerPC 970.

The interview was aimed primarily at filling any lingering gaps in my previous 970 coverage, i.e. corrections, clarifications, and answering some of the questions raised by the articles. So to that end, we talked mostly about the 970's microarchitecture and specifically about the 970's VMX (a.k.a. Altivec) unit, which turns out to be a lot more flexible and robust than what I'd initially described in my articles.

While forward-looking topics were generally excluded from the interview, the article reveals more technical details regarding the PowerPC 970.

 

[nuckinfutz]

Good news on the Altivec clarifications. I'll read the article in it's entirety.

 

[unreg]

optimized compiler

From what I have read here and elsewhere an optimized compiler can speed overall system almost as much as new hardware. This is really good news. As a side benefit, code can be produced without as much hand tweaking for good results.

 

[jamall]

This news will have the execs over at Intel shaking a little more (well, at least sitting up and paying attention). Performance gains from optimisation of the compiler, combined with a shift to the smaller 0.09 micron production process, and perhaps longer term, improvements to the VMX unit and integration of the memory controller, all point to a bright future for PPC computing. Perhaps an extended stay at the top of the pecking order is what Apple needs to win some of that seemingly elusive market share...

 

[ldjessee]

Why not post the whole conversation?

I am not sure why the author at Arstechnica is not posting the whole conversation.

Why edit it and chop it up.

Sure, it makes it easier to digest, but it just leaves too much room for the author's view on what the comments meant. Why not let us read it and let us make that evaluation?

 

[Snowy_River]

I've got to say, the more I read about the 970 since WWDC the more excited I am about its arrival. This chip looks to usher in a new era for Apple, the era that we hoped was here with the G3, until the G4 fiasco crushed those hopes.

"The future's so bright, I have to wear shades..."

 

[MrMacMan]

Omg, I read the original thing when it came out, I know some of you guys can read Swahili, but I can't.

The specifications was like how does the 2ed pipeline link to the 3ed during the fetch/decode sequence.
Me: 'What the hell am I reading?'
Them: 'the vector unit will be processed by second stage ...'
Me: WHAT IS THIS?!?!?!

I mean I skipped like 1/3 of the article.
Anyone who can read thatI take my hat off to you.

 

[Rincewind42]

Re: Why not post the whole conversation?

Originally posted by ldjessee
I am not sure why the author at Arstechnica is not posting the whole conversation.

Why edit it and chop it up.

Sure, it makes it easier to digest, but it just leaves too much room for the author's view on what the comments meant. Why not let us read it and let us make that evaluation?

Probably because there were comments made that they asked him not to print, because there were details that were off topic, or simply because the order of the original conversation made little or no sense. Or all 3 to some degree. The reality is he said he won't post it, so there's no reason to worry about it.

 

[jaedreth]

GCC compiler blues

This is very encouraging news. Not only is the VMX far more flexible than previous thought, it's been made painfully clear that whatever is shipping with the G5's this year is nowhere near optimized for the machine. The compiler itself needs to be optimized for coding for this specific architecture, because this archetecture "breaks the customs" for most industry designs.

That means that as they develop and perfect an optimized version of GCC or something else, then developers can start using this new optimized complier to create optimized code, which will then show even more performance gains out of the same hardware.

So the G5's future looks brighter and brighter.

Can't wait to get mine.

Jaedreth

 

[sparkleytone]

this cleared up some confusion i was having with the FSB speed. the clock multiplier doesnt always have to be 2x. the engineers said there are more multipliers such as 3,4, and 6. This basically means that they don't have to limit processor speed because the FSB can't run at 1.5GHz or whatnot.

 

[ddtlm]

jaedreth:

it's been made painfully clear that whatever is shipping with the G5's this year is nowhere near optimized for the machine

Wowa, this is how wild rumors and crazy expectations get started. They said that GCC wasn't entirely happy with the G5, but that's a lot closer than "nowhere near" optimized.

ldjessee:

Why not let us read it and let us make that evaluation?

The author likely knew that if he just threw the facts out there that some users would get entirely the wrong idea about things.

MrMacman:

Anyone who can read thatI take my hat off to you.

Just take some processor architecture classes from some university, then you can keep your hat on.

 

[Cubeboy]

Re: GCC compiler blues

Originally posted by jaedreth
This is very encouraging news. Not only is the VMX far more flexible than previous thought, it's been made painfully clear that whatever is shipping with the G5's this year is nowhere near optimized for the machine. The compiler itself needs to be optimized for coding for this specific architecture, because this archetecture "breaks the customs" for most industry designs.

That means that as they develop and perfect an optimized version of GCC or something else, then developers can start using this new optimized complier to create optimized code, which will then show even more performance gains out of the same hardware.

So the G5's future looks brighter and brighter.

Can't wait to get mine.

Jaedreth

AIX 5.0 seems to work well with the G5/PPC970 as well, at least for SPEC anyways. The GCC compiler used by Apple was already optimized over standard GCC, hence the correct values for the md file, modified scheduler, and whatever other changes made before WWDC.

"As it turns out, they were in fact wrong before WWDC, but the version of gcc that Apple was using for WWDC had the correct values in it. (I'd imagine that this pre-WWDC obfuscation of the 970's vector latencies was done deliberately, but I forgot to ask him about that."

"The gcc scheduler is not really designed ideally for a processor like the 970 and the Power4 and others, and that's a lot of what the IBM and Apple teams have worked on".

And my favorite quote:
"IBM is not gonna try to compete with Apple's reality distortion field "

EDIT:

It should be noted that GCC 3.3 is not able to schedule for the Pentium 4 at all and has to rely on mcpu=pentium4 which really only tweaks some instruction costs for some very specific problems. Having well scheduled code is critical for nearly all floating point code.

The other thing I've found is that the Pentium 4 doesn't have enough registers for -march=pentium4, since the target machine is a 32 register RISC chip. I'm not sure exactly how much of a performance hit this would cause but a Opteron, which does have enough registers scores much closer to ICC running GCC compiled code, despite having SSE/SSE2 (And we all know ICC is much better at discovering vectorizing opportunities than GCC).

 

[C14ru5]

From the original article
Hannibal: So what are you guys using in the blades, then--the 970 blades? A chipset of your own design?

Peter Sandon: I haven't kept up with what we've announced about that, so I guess I don't know.

Aha! So THAT'S where the alien technology that Steve aquired came to good use

 

[wizard]

Some how I just don't think intel is worried. If we saw Apple quadrupal their market share that might be a differrent story, but right now I can't see Intel worried about a market they don't compete in anyways.

If Apple came up with a switch campaigne that actually worked that might start to worry Intel. But they need more than a switch gimick, they really need to market the positive aspects of their operating system.

I do have to agree the future looks fantastic. Now if they (Apple & IBM) can just move forward with improved 970's to hold the fort until the 980's arrive we will be all set.

Dave

Originally posted by jamall
This news will have the execs over at Intel shaking a little more (well, at least sitting up and paying attention). Performance gains from optimisation of the compiler, combined with a shift to the smaller 0.09 micron production process, and perhaps longer term, improvements to the VMX unit and integration of the memory controller, all point to a bright future for PPC computing. Perhaps an extended stay at the top of the pecking order is what Apple needs to win some of that seemingly elusive market share...

 

[DrugsBunny]

Re: Intel worried

Intel should be worried because they sell CPUs and IBM has produced something cheaper and faster in its first incarnation running code produced by a hacked copy of gcc. Intel has reached the limits of what it can do with the P4 more or less (it can't just keep ramping up the clock-speed forever) while IBM has only just started. Itanium2 is DOA for the desktop. The days of Mac users and the PowerPC faithful waiting for anaemic increments to G4 clockspeeds on a bandwidth-starved bus from Motorola are over. Folks, welcome to flavor country.

 

[8thDegreeSavage]

*In best Homer voice: "I'm in flavor country"

 

[Sun Baked]

Notice, no real response to the PPC 970 White Box comment.

For the PPC White Box crowd the addition of PPC 970 machines would be a welcome addition to Mai Logic and Pegasos's G3/G4 motherboards (and others).

More desktop PPC machines does help Apple out, and will keep IBM focused on improving the product. It's been a long road to recovery for this class of machine since MS killed Windows for PPC.

Should be interesting to see if this keeps the Amiga on PPC alive or the bail and head over to x86. But Amiga has been working with MS lately.

 

[websterphreaky]

What G5's, you may not be seeing any G5's!!

Wonder why MacRumors isn't reporting that IBM is having REAL BIG problems at the Fishkill plant on the 970 production line??

You may not be getting those G5's afterall, or greatly delayed.

[mod. edit - Next time search the forums before you try to troll. That story was posted last week when it first came out.]

 

[GetSome681]

Re: What G5's, you may not be seeing any G5's!!

Originally posted by websterphreaky
Wonder why MacRumors isn't reporting that IBM is having REAL BIG problems at the Fishkill plant on the 970 production line??

You may not be getting those G5's afterall, or greatly delayed.

Wow, too bad that was posted here like a week ago. And for you people that don't understand business speak, that article has nothing to do with production problems as in you not being able to get your G5s, it has to do with the plant losing money, since they aren't producing enough with it to turn a profit on their investment. This will change though. Keep your panties on.

 

[MrMacMan]

Originally posted by ddtlm

MrMacman:


Just take some processor architecture classes from some university, then you can keep your hat on.

Hey, I think that can wait a few years.

I'm not exactally in college, I don't think learning how processors work will help me...

And if the G5 comes out on schedule we will all be good.

 

[ddtlm]

DrugsBunny:

Optimism aside, I see no reason to believe any of what you claim. People predict the death of x86 just about as often as they predict the death of Apple... has either happened?

IBM has produced something cheaper and faster in its first incarnation

"Power4" has a "4" in its name because the generation before it had a "3" in its name.

running code produced by a hacked copy of gcc

Great, now everyone is gona blame GCC for every failing of Apple just like Moto was/is blamed. Intel gets higher benchmark score? No problem, just blame GCC!

 

[DrugsBunny]

Note that I never predicted the death of the x86. I predict that Intel will no longer be able to get much out of the P4 core. i.e. using a thinner process and clocking it faster only gets you so far, then you need to use a new core design.

As far as first incarnation is concerned, the 4 (as in Power4) designation is irrelevant. Recall that the G4 first appeared as a 500Mhz chip (I think). That's what I call a first incarnation. Subsequent revisions (7400 vs 7450 vs 7457 each at various clock speeds) increased the clock speed and had minor differences internally. In that sense, the PPC970 is the "first version" of that class of chip.

Again, at no stage did I blaim all of Apple's failings on gcc or mot and had Apple run the spec tests using Intel's own compiler with the P4, the SPECint tests would have shown the P4 winning by an even larger margin. The PPC970 is large beast of a processor and can handle a lot of in-flight instructions (216?) at a time. The proviso is that there are constraints on what instructions can be issued together in a given group. gcc is designed on a model of a processor where there's a pipeline and it reorders stuff on a per-instruction level to avoid stalls. It's totally ignorant of this whole group issue/dispatch business. The hack to make it work is that you fakeout what might be an equivalent pipeline and make some educated guesses on the code that should be generated for the 970. It works but is not the ideal situation.

 

[Snowy_River]

Originally posted by DrugsBunny
Note that I never predicted the death of the x86. I predict that Intel will no longer be able to get much out of the P4 core. i.e. using a thinner process and clocking it faster only gets you so far, then you need to use a new core design.

This is actually a very good point. I've read that Intel is running up against a wall, performance wise. That's not to say that they can't keep clocking their chips faster. Quite the contrary, they can, but they're reaching into the territory of deminishing returns.

I read an article that said that the x86 architecture at this point only gains about 2% processor performance for a 10% increase in clock speed. Now, doing some quick math, that means that to get a P4 that out-performs a 3GHz P4 by 10%, you'll need a 4.6GHz P4. Yes, that's right, a 60% increase in clock speed. This is the wall that Intel is hitting.

Now, the downside of this is that there will still be those who will say "I've got a 4.6GHz P4, and it can smoke your puny 2.5GHz G5." The MHz (or should that be GHz?) myth is still out there, and probably will be for some time. However, with entrants like the PPC970 on the field, things may start turning around a little faster than they have been.

 

[Analog Kid]

Originally posted by MrMacman
Hey, I think that can wait a few years.

I'm not exactally in college, I don't think learning how processors work will help me...

And if the G5 comes out on schedule we will all be good.

To be honest, you probably don't need to be in college to understand a lot of the basics. Not that you don't have better ways to spend your time, but you could probably get a good understanding of it by reading a book on architecture design if you're interested.

A lot of it is just visualization skills and learning the jargon.

Maybe by the time you figure it out they'll be hiring engineers again...

 

[Analog Kid]

Originally posted by Snowy_River
This is actually a very good point. I've read that Intel is running up against a wall, performance wise. That's not to say that they can't keep clocking their chips faster. Quite the contrary, they can, but they're reaching into the territory of deminishing returns.

I'd be very surprised if Intel doesn't know exactly where the P4 will hit the wall and doesn't have a plan to push past it.

And I doubt they're relying on the Itanic to bail them out. Ironically though, it might be AMD that keeps the x86 ISA alive while Intel re-tools.

That said, I do like the statement that IBM is gearing to ramp up just as Intel is tapering off. It's a wonderfully optimistic feeling!

We have a roadmap, folks!
There's a future!

It's gonna be a fun couple years!



Oh, and Mot is responsible for all of Apple's problems... And then IBM arrived on a brilliant white charger, armor gleaming in the sun...