PowerPC 970 Presentation PDF

The link has already been disabled. You may want to post the content to a mirror.

Rocketman

The link still worked for me at 9AM EST!

You can go here and grab it if you need to. At least until they block it for excessive bandwidth use.

Originally posted by Telomar
You can go here and grab it if you need to. At least until they block it for excessive bandwidth use.

Click to expand...

I've put it up on my iDisk to, it can also be got from here if Telomars stops working.

Good Reading but it sure seems it could be a while before the 970 power mac will be in my house!~( I'll wait till at least rev. 2 before I buy.

Thanks for posting that,

Joe

Re: PowerPC 970 Presentation PDF

Originally posted by Macrumors
Volume shipments of the processor are expected in late 2003.

Click to expand...

So what exactly constitutes volume shipment?

Could we see a smaller volume initially going to Apple for PowerMacs and or Xserve, and then, a volume shipment going into the other lineups and Linux based systems towards the end of 2003? In effect the current top-of-the-line G4 would be for the consumer lineup and PowerBooks come January?

The sooner the better. Could 'late 2003' be Q3, or is it gunna be in Q4?

Double floating point?

I don't see anything about double floating point within their VMX (the 162 instruction SIMD) in the document. That's not particularly good.

However, it does say it does have hardware double floating point support, just hope it doesn't come from the G4 which is from 60xe o).

I am confused...according to this, when would it be likely to see this processor integrated into a new line or revision of Powerbooks? Also, how much faster will this processor really be than the current G4's (if you were comparing both processors at the same Mhz, theoretically)? And is another Powerbook G4 speedup likely before this processor is integrated? If so, can anyone make an educated guess as to about when?

-Cameron

Originally posted by cjerens
I am confused...according to this, when would it be likely to see this processor integrated into a new line or revision of Powerbooks? Also, how much faster will this processor really be than the current G4's (if you were comparing both processors at the same Mhz, theoretically)? And is another Powerbook G4 speedup likely before this processor is integrated? If so, can anyone make an educated guess as to about when?

-Cameron

Click to expand...

I wouldn't even be thinking about the PowerBook, these will have to make it ito the PowerMac first.

Re: Double floating point?

Originally posted by MacCoaster
However, it does say it does have hardware double floating point support, just hope it doesn't come from the G4 which is from 60xe o).

Click to expand...

No, it should be much better... according to c't a 1GHz G4 system had 187 SPECfps, while the PPC970 slides indicate 1000+ SPECfps.

Originally posted by edesignuk
I wouldn't even be thinking about the PowerBook, these will have to make it ito the PowerMac first.

Click to expand...

And maybe the XServe before that!

As to the actual performance compared to current G4s (cjerens), I don't think anybody knows for sure. IBM gave some estimates for SPEC benchmark scores, but it's hard to extrapolate too much from that.

Can anyone give an educated estimate of how much faster this 1.8Ghz 970 will be when compared to Motorolas yet to come out 1.8Ghz G4?? What sort of percentages are we talking about? I dont want wishfull thinking figures. I want a realistic estimate.

Any ideas???

If Apple can get these chips at a good price, I think they would put it in a Powermac first. Apples servers arent exactly what people want when they think of Apple, it would be a waste to have them only in the Xserve next year.
Also, if they officially say end of 2003 that means that Apple will probably have these out by the start of summer or even in the $3,000+ model in the start of the year. I think thats what they will do, put it in the $3,000+ model with todays motherboard and wait till the end of the year or the start of next to give us the ApplePI motherboard and show off its true power then.
Has Apple ever given us a new technology/CPU/Motherboard that was all perfect when it first came out? It usually gets as good as it should have been in its second or third revision. And going by their track record, I expect nothing different this time. Either way, we will see quite a speed boost all through next year, if all goes well.

Welll...

Originally posted by Falleron
Can anyone give an educated estimate of how much faster this 1.8Ghz 970 will be when compared to Motorolas yet to come out 1.8Ghz G4?? What sort of percentages are we talking about? I dont want wishfull thinking figures. I want a realistic estimate.

Any ideas???

Click to expand...

...
Assume that SPEC scores scale linearly with clock frequency (which they don't, but it's easier to compare that way). A 1.8GHz G4+ would get 1.8x180 (roughly), or 344 on SPECFP. The 970 gets about 1050. The G4 would get about 540 on SPECINT, the 970 gets about 900 (note: SPEC is double precision heavy, and doesn't use Altivec). Why?
1) Memory bus bandwidth: 1.3GBps vs. 6.4GBps (about a 5x improvement)
2) Instruction dispatches per cycle: 3 (iirc) vs. 5 (close to a 2x improvement)
3) Floating point units: 1 vs 2 (2x improvement)
4) Out of order execution: none vs lots
5) Compiler: GCC vs IBM-übercompiler (whatever it's called)
6) 16 instructions "in flight" at one time vs. 40 groups of 5 (200)
7) the 970 has much more sophisticated branch prediction

I would expect the 970 to get close to double the floating point performance of a G4+ at the same clock frequency. Singly precision integer performance should be closer to equal. The 970 will be significantly ahead on double precision integer. For Altivec (VMX) code, I think the G4+ would have an edge over the 970, except that it doesn't have enough memory bandwidth (i.e. if you gave a 1.8GHz G4+ 6.4GBps of memory bandwidth, I think it would win at Altivec code).

I think it will average about 1.5-2x faster than a G4+ at the same clock frequency. I would also guess that it should be roughly equivalent in performance to Pentium 4s of the same time period (I'm guessing 3.6GHz, .09 micron, hyperthreading, 667MHz bus, >512k cache).

Re: Welll...

Originally posted by Catfish_Man
...
Assume that SPEC scores scale linearly with clock frequency (which they don't, but it's easier to compare that way). A 1.8GHz G4+ would get 1.8x180 (roughly), or 344 on SPECFP. The 970 gets about 1050. The G4 would get about 540 on SPECINT, the 970 gets about 900 (note: SPEC is double precision heavy, and doesn't use Altivec). Why?
1) Memory bus bandwidth: 1.3GBps vs. 6.4GBps (about a 5x improvement)
2) Instruction dispatches per cycle: 3 (iirc) vs. 5 (close to a 2x improvement)
3) Floating point units: 1 vs 2 (2x improvement)
4) Out of order execution: none vs lots
5) Compiler: GCC vs IBM-übercompiler (whatever it's called)
6) 16 instructions "in flight" at one time vs. 40 groups of 5 (200)
7) the 970 has much more sophisticated branch prediction

I would expect the 970 to get close to double the floating point performance of a G4+ at the same clock frequency. Singly precision integer performance should be closer to equal. The 970 will be significantly ahead on double precision integer. For Altivec (VMX) code, I think the G4+ would have an edge over the 970, except that it doesn't have enough memory bandwidth (i.e. if you gave a 1.8GHz G4+ 6.4GBps of memory bandwidth, I think it would win at Altivec code).

I think it will average about 1.5-2x faster than a G4+ at the same clock frequency. I would also guess that it should be roughly equivalent in performance to Pentium 4s of the same time period (I'm guessing 3.6GHz, .09 micron, hyperthreading, 667MHz bus, >512k cache).

Click to expand...

Thanks, I'd say that was educated. I know I am getting ahead of myself but it would be great if IBM could come out with a 2Ghz 970 just so that they were ahead of the PC world for once (in performance anyway).

(note: SPEC is double precision heavy, and doesn't use Altivec). Why?

Click to expand...

Altivec is single precision, at least on the G4. it executes once every clock cycle. SSE2 on the Pentium 4 takes 2 clock cycles to execute but it's double precision. That could be the reason, even with altivec, it's not a fantastic cpu for SPEC benchmarks but it screams at a few realworld tasks.

Can't wait for the PPC970 based powermacs to come out. No matter what intel or AMD come out with, they're not going to running 6Ghz Pentium 4's or 4Ghz Athlons so it could put apple back on top again and they run those snail adverts again.

Re: Re: PowerPC 970 Presentation PDF

Originally posted by pgwalsh

So what exactly constitutes volume shipment?

Click to expand...

Volume shipment is a term used to indicate when the vendor can ship enough parts that the customer is actually able to build and ship their product in the required volume.

In the case of Apple (or any other computer builder for that matter) there is usually a period of inventory buildup where a new component (eg PPC 970) is stockpiled enough to allow the company to meet predicted orders. They will not normally try to ship products when components are trickling in because it is impossible to tell the customer when the product will be shipped.

You often find that companies will announce products when they start to receive volume shipments of components, and then start to deliver 2 - 4 weeks later once they are comfortable that they can ship consistently.

Note...

Originally posted by barkmonster
Altivec is single precision, at least on the G4. it executes once every clock cycle. SSE2 on the Pentium 4 takes 2 clock cycles to execute but it's double precision. That could be the reason, even with altivec, it's not a fantastic cpu for SPEC benchmarks but it screams at a few realworld tasks.

Can't wait for the PPC970 based powermacs to come out. No matter what intel or AMD come out with, they're not going to running 6Ghz Pentium 4's or 4Ghz Athlons so it could put apple back on top again and they run those snail adverts again.

Click to expand...

...that the why isn't in the (). The why is "why does the 970 score so much higher than the G4+?". I answered my own question below it. Sorry about the misunderstanding.

Re: Welll...

Originally posted by Catfish_Man
7) the 970 has much more sophisticated branch prediction
For Altivec (VMX) code, I think the G4+ would have an edge over the 970, except that it doesn't have enough memory bandwidth

Click to expand...

The branch prediction is better, but that might be offset by the longer pipelines compared to the G4. (Then again I don't know when branches are decided in the 970.)

Why do you think the G4 Altivec unit has the advantage? The 970 looks to be pretty similar on paper, with 2 integer, 1 fp and 1 permute unit.

Originally posted by barkmonster
Can't wait for the PPC970 based powermacs to come out. No matter what intel or AMD come out with, they're not going to running 6Ghz Pentium 4's or 4Ghz Athlons

Click to expand...

Interestingly, the current 2.8GHz P4 gets roughly the same SPEC scores as the estimates of the 1.8GHz 970. (See http://www.spec.org/osg/cpu2000/results/ ) I know SPECs are just one (set of) benchmarks, the machine was probably stacked, and you're right that there ain't gonna be no 6GHz P4 coming out, but those current P4 numbers are still a bit surprising!

Interestingly, the current 2.8GHz P4 gets roughly the same SPEC scores as the estimates of the 1.8GHz 970. (See http://www.spec.org/osg/cpu2000/results/ ) I know SPECs are just one (set of) benchmarks, the machine was probably stacked, and you're right that there ain't gonna be no 6GHz P4 coming out, but those current P4 numbers are still a bit surprising!

Click to expand...

Thanks for the link!

Here's a few of the results, I've put the motherboard manufacturer and RAM type in brackets

SPECfp_base2000

AMD Athlon 2800XP+ (2.25 Ghz, ASUS, 333Mhz DDR) : 782
Intel 2.8Ghz P4 (Intel, 1066Mhz RDRAM) : 1032
Intel 1.8Ghz P4 (Intel, 800Mhz RDRAM) : 699
IBM 1.8Ghz PPC 970 : 1051

SPECint_base2000

AMD Athlon 2800XP+ (2.25 Ghz, ASUS, 333Mhz DDR) : 898
Intel 2.8Ghz P4 (Intel, 1066Mhz RDRAM) : 1034
Intel 1.8Ghz P4 (Intel, 800Mhz RDRAM) : 633
IBM 1.8Ghz PPC 970 : 937

One thing that's got me thinking, on the same page as the SPEC scores in the PDF, there's these GFLOP ratings :

Peak scalar GFLOPS = 7.2

Peak SIMD GFLOPS = 14.4

I know in altivec fractal benchmarks, turning off altivec dropped the G4s performance from several GFLOPS to only a few hundred thousand FLOPS, it looks like even without Altivec these chips are going to be significantly faster than the G4. I'm assuming scaler means using only the RAW CPU power because SIMD is refering to the VMX/Altivec unit.

Oops! got the FP and INT figures for the PPC 970 backwards, thanks for the tip avkills

Numbers backwards!

Barkmonster, you have the PowerPC970 numbers switched.

The spec2000int is 937

The spec200fp is 1051

This processor looks like it is going to kick some serious ass. The fact that it beats a 2.8Ghz P4 in floats is awesome, one serious 3d rendering monster. And besides, the clock frequencies are subject to change, and since it only dissipates 42W, I could see IBM cranking it up to 2+ Ghz, which would still work in a tower with good cooling.

-mark

Wow, this new IBM chips sounds really great. Does that mean that when it comes out the finder on iMacs and iBooks will be as fast as it is on a Powermac 1.25ghz? Because if I have to spend over $3,000 on a computer that will give me the finder speed of OS 9, it would be kind of a let down.