Analysis of the 970

[rice_web]

The 970 is already more than a prediction by most of the MacRumors gang. So, I'd like to provide a comparison of the Apollo and our potential G5.

G4 - Motorola's Flagship Chip
Motorola has only been able to produce a 1.25GHz (that is apparently available). But, the G4 has a few advantages going for it. The G4 has Altivec, only sports seven pipeline stages, and supports two megabytes of L3 cache.

However great those improvements may be, however, they are not enough to fight off a pathetic 166MHz system bus, the P4s SSE2, and dated manufacturing method.

970 - The Chip We'd Like To Call The G5
The 970 from IBM is an excellent chip, and if it were released by the end of the month, would still battle fiercly with the P4 and Athlon. However, the 970 probably won't find itself in production until the second half of 2003, and likely will not find any applications until early 2004.

Still, I like to be optimistic. If the chip is initially available at 1.4-1.8GHz, Mac fans would gain a significant performance hike over the G4--even if it is unable to beat Intel and AMD's offerings.

7455
- Up to 1.5GHz (overclocked)
- 166MHz System Bus
- SDRAM Support
- 2MB L3 Cache (1/4 Processor Clock)
- 256K L2 Cache (Full Processor Clock)
- Altivec
- 32-bit processing
- 180nm manufacturing

970
- Up to 1.8GHz (IBM claim)
- 900MHz System Bus (450 x 2)
- DDR RAM Support
- 512K L2 Cache (Full Processor clock)
- Altivec
- 64-bit processing
- 130nm manufacturing (90nm rumored)

 

[Telomar]

Re: Analysis of the 970

The only real benefit the G4 has over the PPC970 is the L3 cache. The PPC970 really doesn't need it half as much though.

Originally posted by rice_web
However great those improvements may be they are not enough to fight off the P4s SSE2.

I would take Altivec over SSE2 10/10 times and I imagine most other people would as well.

 

[Nipsy]

If a 970 comes out in a Mac with the rumored 6.4GB/s bandwidth, it will be faster in day to day use than the SPEC predictions lead you to think.

Even a G4 with greatly increased I/O bandwidth would outclass the G4s of today.

SPEC scores (int and fp) test the processor, but don't stress the rest of the system, so even if the x86 processor(s) score(s) slightly higher in SPEC (int and fp) when flagships are tested next year, the actual speed for an end user may favor the Mac heavily. Unless we see something huge in the PC mobo/chipset/architecture world...

P.S. We don't yet know what amount of L3 the 970 will be delivered with...

 

[ddtlm]

Telomar:

I would take Altivec over SSE2 10/10 times and I imagine most other people would as well.

I'm sure Apple appreciates your loyalty, but when you want to do double precision floating point SSE2 is clearly the choice, since AltiVec can't do it at all. That should be at least 1/10 times.

Nipsy:

Actually saying that SPEC doesn't use memory bandwidth, or that it only stresses the processor is silly. Notice how much the Power4 leads the PPC-970 in SPEC floating point? That has everything to do with memory bandwidth (including L3) since the execution cores are exactly the same (except where the PPC-970 has added things like AltiVec).

SPEC scores (int and fp) test the processor, but don't stress the rest of the system, so even if the x86 processor(s) score(s) slightly higher in SPEC (int and fp) when flagships are tested next year, the actual speed for an end user may favor the Mac heavily

What do you base this on? Since obviously SPEC stresses both memory and CPU (based on PPC-970 vs Power4), all that is left are things like video cards and hard disks, and those are the same for Macs and PCs.

It is true that SPEC does not reflect Photoshop or RC5 scores, and I expect those will favor Macs as much as ever, probably even more so than ever before.

 

[MacBandit]

Re: Analysis of the 970

Originally posted by rice_web

7455
- Up to 1.5GHz (overclocked)
- 166MHz System Bus
- SDRAM Support
- 2MB L3 Cache (1/4 Processor Clock)
- 256K L2 Cache (Full Processor Clock)
- Altivec
- 32-bit processing
- 180nm manufacturing

Just when you think that people have gotten the point about overclocking here comes more of it.

Overclocking is a term to describe what an end user does to a processor and can not and has never been used in the sense of a manufacturor. The new PPC7455a chips from Motorola that Apple is selling at 1250Mhz are tested and thusly stamped with a rating of 1250Mhz.

 

[Nipsy]

Re: Re: Analysis of the 970

Originally posted by MacBandit


Just when you think that people have gotten the point about overclocking here comes more of it.

Overclocking is a term to describe what an end user does to a processor and can not and has never been used in the sense of a manufacturor. The new PPC7455a chips from Motorola that Apple is selling at 1250Mhz are tested and thusly stamped with a rating of 1250Mhz.

I think he was pointing out that end users are indeed overclocking the 1250 MHz chip you reference, to 1.5GHz, as he states.

 

[MacBandit]

Re: Re: Re: Analysis of the 970

Originally posted by Nipsy


I think he was pointing out that end users are indeed overclocking the 1250 MHz chip you reference, to 1.5GHz, as he states.

That doesn't seem to be what he's inferring there. It just looks to me that he's comparing future specs of a PPC7455 to a PPC970.

 

[Sherman]

And then there's also the insanely fast system bus.

Lest we forget a 450Mhz DDR system bus, and hopefully, full DDR RAM support?

Even if the 970 ran at the same clock speed, it would still probably blow the doors off of our measly 1.25 Ghz G4. The crappy system bus we have and the lackluster DDR support is really cramping the G4's style.

And these 2004 predictions are ridiculous. Apple has been planning to use the 970 for quite some time now. think about it. Why hasn't apple announced ANYTHING new in the hardware department. Nothing. Nada. Zilch.

Spidey sense, TINGLING!

 

[Nipsy]

Originally posted by ddtlm
Telomar:

Nipsy:

Actually saying that SPEC doesn't use memory bandwidth, or that it only stresses the processor is silly. Notice how much the Power4 leads the PPC-970 in SPEC floating point? That has everything to do with memory bandwidth (including L3) since the execution cores are exactly the same (except where the PPC-970 has added things like AltiVec).


What do you base this on? Since obviously SPEC stresses both memory and CPU (based on PPC-970 vs Power4), all that is left are things like video cards and hard disks, and those are the same for Macs and PCs.

It is true that SPEC does not reflect Photoshop or RC5 scores, and I expect those will favor Macs as much as ever, probably even more so than ever before.

What I'm talking about is memory, PCI, AGP, northbridge, etc. bandwidth outside of cache. Cache is highly integrated with the processor, so in int and fp scoring, more cache helps, however, int anf fp routines rarely leave the bounds of L1, L2, & L3, where memory bandwith, system I/O, and even paging to disk can all affect relative performance.

Notice how SPEC scores remain virtually constant as RAM/storage is added?

A slightly more balanced way of measuring full system perfromance in TPM, where a database, the processor, and the RAM are all stressed.

The current G4 is starved for sytem bandwidth, meaning that when you're waiting, your waiting for the disk to write to memory to feed instructions to the processor, much more often than you're waiting for the processor to execute instructions.

Our current architecture uses many small pipes feeding a medium valve. The bottleneck is more often the pipes than the valve. Hopefully the 970 and associated MoBo will change that to many medium/large pipes feeding a large valve, meaning that the valve will be constantly supplied.

 

[Telomar]

I have a hunch the first version of the PPC 970 won't have a L3 cache controller in order to reduce pin count. Might be there but I imagine it would have merited some mention.

Originally posted by ddtlm
I'm sure Apple appreciates your loyalty, but when you want to do double precision floating point SSE2 is clearly the choice, since AltiVec can't do it at all. That should be at least 1/10 times.

I knew somebody would raise that. It is actually one of the few weaknesses in Altivec and iwhen I was writing the post I had mentioned it then I cut it.

Given the fact altivec uses separate registers and we are talking about a 64-bit chip I would still take altivec 10/10 times. Overall it is just a better package...unless you count optimising and programming for it. There is a good reason people that do a lot of work in doubles often use 64 bit workstations already instead of Pentium IVs with SSE2.

 

[Postal]

Ah, but the whole reason PowerMacs and PowerBooks come with L3 cache is because there isn't enough system bandwidth for the way it would work otherwise (i.e. just sending the data through). The PPC 970 may not need L3 for awhile (if at all) when there's a 6.4 GB/sec interface between the CPU and northbridge chip.

 

[Bradcoe]

Less $$ too

No L3 means apple COULD charge much less than with L3 cache. That stuffs expensive in 2 and 3MB chunks.
Heres another point I haven't heard anyone talk about. DDR RAM prices. I don't want to think of how much 900Mhz DDR (actual 450) is going to cost. And if Apple is retarded, I REALLY don't want to think about saving $ by putting slower RAM, or even worse, a slower memory controller made by Apple into these machines. I'm slightly nervous about pricing. I see lots of reasons for the 970 to be a reason to make Mac's less money. However, I see Apple as having lots of reasons to make the 970 responsible for equally high or even higher priced macs.

 

[benixau]

Apple will follow its usual strategy.

1st move: 1st gen processors, big numbers, big price
2nd move: slight to moderate price drop
3rd move: 2nd gen processors, bigger numbers, slight price rise
4th move: sligh to moderate price drop

etc etc

This is what they have always done and why would they stop now??

 

[Catfish_Man]

Re: Less $$ too

Originally posted by Bradcoe
No L3 means apple COULD charge much less than with L3 cache. That stuffs expensive in 2 and 3MB chunks.
Heres another point I haven't heard anyone talk about. DDR RAM prices. I don't want to think of how much 900Mhz DDR (actual 450) is going to cost. And if Apple is retarded, I REALLY don't want to think about saving $ by putting slower RAM, or even worse, a slower memory controller made by Apple into these machines. I'm slightly nervous about pricing. I see lots of reasons for the 970 to be a reason to make Mac's less money. However, I see Apple as having lots of reasons to make the 970 responsible for equally high or even higher priced macs.

Actually, it's an 800MHz effective bus (1 bit in 9 is overhead). That matches up quite nicely with dual channel DDRII-400, which (coincidentally ) is going to be the current high end ram when the 970 comes out.

 

[DaveGee]

Originally posted by benixau
Apple will follow its usual strategy.

1st move: 1st gen processors, big numbers, big price
2nd move: slight to moderate price drop
3rd move: 2nd gen processors, bigger numbers, slight price rise
4th move: sligh to moderate price drop

etc etc

This is what they have always done and why would they stop now??

Well up till now we have been stuck with MOT and their (hmmm how do I say this) less then stellar track record of bunny hop progress. The PRO line and the consumer line has been pretty tight over these past many years. Much of the time video card, and a second CPU are what set apart the pro line. Soon we will have the IBM 97x family (yes more are on the way) of CPUs. They are a HUGE leap from what we have now not only in speed but in other areas as well.

Apple has to make a choice.

1 - Hold back on the high end so the low end doesn't look SO BAD.
2 - Balls to the walls the high end and have the low end move up to the 970 CPUs as well. Even with this laptops could still be a problem.

Both have probems of their own...

Option 1 is just a stop gap and doesn't really address the problem. In the end the consumer line has to advance.

Option 2 could mean (guessing) higher prices or lower margine. Neither of which Apple can really afford and in this economy faster and lowercost have to go hand in hand otherwise the boxes will NOT sell even if they are 'da goods'.

Laptops also pose the same problem. Until we hear that 'yes the 970 will work well in a laptop' (from an official source) Apple might be forced into option #1...

Remember the consumer line sells so well (out sells the PRO line by a HUGE margin) all because they are a 'good value' when compared to the PRO line.

Who here hasn't said: You know the top of the line iMac is a pretty good box given the feature set and for the most part is pretty darn close (in speed) to the bottom end tower.

Once the PRO goes 970 what will people say about the low end?

Marketting sucks...

Dave

 

[rice_web]

That doesn't seem to be what he's inferring there. It just looks to me that he's comparing future specs of a PPC7455 to a PPC970.

I was referring to the overclocked chips, which represents a possibility max. clock speed for the G4 on this manufacturing process.

P.S. We don't yet know what amount of L3 the 970 will be delivered with...

IBM has stated that the 970 will not support an L3 cache.

I would take Altivec over SSE2 10/10 times and I imagine most other people would as well.

As would I. I wasn't actually giving SSE2 the upper hand. SSE2 is nice (very nice, actually), but it is no match for Altivec (on certain tasks). My reference was to the P4's amazing clock speed WITH SSE2; a combination of both (though I didn't make that very clear)

 

[ddtlm]

Nipsy:

I hate to keep dragging off into arguements about SPEC, but people plainly do not get it.

What I'm talking about is memory, PCI, AGP, northbridge, etc. bandwidth outside of cache. Cache is highly integrated with the processor, so in int and fp scoring, more cache helps, however, int anf fp routines rarely leave the bounds of L1, L2, & L3, where memory bandwith, system I/O, and even paging to disk can all affect relative performance.

How can you conclude this after I pointed out the performance edge given to the Power4 by it's large L3? Do you think that data goes nowhere on the PPC-970? If it's not in the 32 meg L3, it's in RAM, and the reason why the PPC-970 looses is because the RAM is slower than L3... hence we can conclude that memory performance matters.

You might direct your attention to this link:

http://www.spec.org/osg/cpu2000/docs/readme1st.html#ovw

Q1, Q3, Q5, Q6 and Q11 are informative. Q9 also lists 256 megs of RAM as a requirement, which strongly suggests that that much RAM is used... which means that the performance of said RAM matters.

Notice how SPEC scores remain virtually constant as RAM/storage is added?

This link should explain that very clearly:

http://www.spec.org/osg/cpu2000/docs/system_requirements.html

A slightly more balanced way of measuring full system perfromance in TPM, where a database, the processor, and the RAM are all stressed.

Only if you are running a server does a server benchmark such as TPM matter.

The current G4 is starved for sytem bandwidth, meaning that when you're waiting, your waiting for the disk to write to memory to feed instructions to the processor, much more often than you're waiting for the processor to execute instructions.

Yes, people desperately want to blame the performance problems of the G4 on it's memory, but really that isn't the case. It's just got a crappy scalar core, lacking things as basic as out-of-order execution and hardware prefetech (although there is a manual prefect in AltiVec... not as good as the auto-pretect in modern x86 chips). As I've said again and again, I have an Athlon on PC133 memory that performs a hell of a lot better than my G4-800-DP.

 

[rice_web]

Yes, people desperately want to blame the performance problems of the G4 on it's memory, but really that isn't the case. It's just got a crappy scalar core, lacking things as basic as out-of-order execution and hardware prefetech (although there is a manual prefect in AltiVec... not as good as the auto-pretect in modern x86 chips). As I've said again and again, I have an Athlon on PC133 memory that performs a hell of a lot better than my G4-800-DP.

This is one reason I love the 750FX. It's architecture is amazing. Thankfully, the 970 promises an architecture that parallels the 750FX, simply adding one pipeline stage (in part to allow 64-bit processing).

 

[rice_web]

How can you conclude this after I pointed out the performance edge given to the Power4 by it's large L3? Do you think that data goes nowhere on the PPC-970? If it's not in the 32 meg L3, it's in RAM, and the reason why the PPC-970 looses is because the RAM is slower than L3... hence we can conclude that memory performance matters.

I'd like to make a comment on the bus speed and memory interface of the 970....

If DDR memory is available at 900MHz (which it currently isn't) by the time we see--if--this CPU in an Apple computer, 900MHz DDR memory would nearly outpace the speed of a L3 cache, despite a greater distance for the information to travel (all the way to the system memory rather than a few millimeters to a cache).

Granted, 900MHz DDR memory will only find itself in video cards, and even then probably not for at least a year. Plus, it'd probably cost the equivalent of a kidney on the black market.

So, we'll probably see DDR400 or DDR533 by the time--if--the 970 winds up in Apple computers, which would still perform extremely quickly (roughly 1/2 the speed of an L3 cache).

 

[Catfish_Man]

I take it...

Originally posted by rice_web


I'd like to make a comment on the bus speed and memory interface of the 970....

If DDR memory is available at 900MHz (which it currently isn't) by the time we see--if--this CPU in an Apple computer, 900MHz DDR memory would nearly outpace the speed of a L3 cache, despite a greater distance for the information to travel (all the way to the system memory rather than a few millimeters to a cache).

Granted, 900MHz DDR memory will only find itself in video cards, and even then probably not for at least a year. Plus, it'd probably cost the equivalent of a kidney on the black market.

So, we'll probably see DDR400 or DDR533 by the time--if--the 970 winds up in Apple computers, which would still perform extremely quickly (roughly 1/2 the speed of an L3 cache).

...that you didn't see my post about dual channel DDRII? That fills up the 800MHz that's useable on the bus quite nicely.

 

[rice_web]

...that you didn't see my post about dual channel DDRII? That fills up the 800MHz that's useable on the bus quite nicely.

I'm amazed that I missed that post. You had a good point before I did.

Whatever the case, the 970 doesn't need an L3 cache to perform well (although, an L3 cache helps tremendously). Our 512K of L2 cache should be enough with the memory bandwidth at our disposal.

 

[ddtlm]

rice_web:

This is one reason I love the 750FX. It's architecture is amazing. Thankfully, the 970 promises an architecture that parallels the 750FX, simply adding one pipeline stage (in part to allow 64-bit processing).

Of course the 750fx lacks hardware pretech and out of order execution (AFAIK) as well. You seem to confuse the efficientcy of a simple design with a good design... I mean of course a short pipeline-CPU is efficient.

If DDR memory is available at 900MHz (which it currently isn't) by the time we see--if--this CPU in an Apple computer, 900MHz DDR memory would nearly outpace the speed of a L3 cache, despite a greater distance for the information to travel (all the way to the system memory rather than a few millimeters to a cache).

You have stumbled onto another "megahertz myth" situation. The situation is complex, but anyway clock speed does not equal performance for memory either. For one thing, the SRAM L3 on G4's is better than SDRAM memory because the latency is enherently lower. Also, as "faster" types of RAM are used, you cannot assume that the latency is beting reduced. For example, the 533mhz FSB on a P4 has more-or-less the same random-access latency as a 133mhz FSB on P3 or G4 because in both cases the base clock speed (133mhz) is the same! DDR-II will send data a hell of a lot faster, but AFAIK the latency is not being reduced. Now all of this is very very complicated and I can't pretend to be qualified to lecture on the subject. It is further complicated by the ability of hardware prefetch in both processors and chipsets to mask the enderlying latency by guessing the next memory access with some random chance of getting it right, so the actual turn around time seems a lot lower from the point of view of the CPU.

Complex stuff. Tread carefully.

 

[MacCoaster]

Originally posted by rice_web
If DDR memory is available at 900MHz (which it currently isn't) by the time we see--if--this CPU in an Apple computer, 900MHz DDR memory would nearly outpace the speed of a L3 cache, despite a greater distance for the information to travel (all the way to the system memory rather than a few millimeters to a cache).

Granted, 900MHz DDR memory will only find itself in video cards, and even then probably not for at least a year. Plus, it'd probably cost the equivalent of a kidney on the black market.

So, we'll probably see DDR400 or DDR533 by the time--if--the 970 winds up in Apple computers, which would still perform extremely quickly (roughly 1/2 the speed of an L3 cache).

I was told that the IBM PowerPC 970 had dual pumped 450MHz DDR channels. Therefore rendering the 900MHz DDR RAM useless. You'd have 1:1 ratio of 450MHz DDR for the two channels for a combined performance of 900MHz.

That's how the current nForce chipsets work, at least.

If anyone has any information that proves me wrong, please do provide.

 

[ddtlm]

MacCoaster:

I'm a little confused what you are saying, but the PPC-970 does support two 32-bit channels as a "FSB" and each clocks at 900mhz. This is a little like how Rambus memory clocks at obscene rates, but only has 8-bit channels.

In any case, there is overhead on the FSB so that the actual rate at which data arrives is 800mhz, not 900mhz (1 bit in 9 is overhead).

This could be paired with dual DDR-400 channels, but I don't know if Apple will go the road... right now DDR-400 isn't even finalized and most all of the boards using it have problems with stability. Notably the nForce2 does support dual channels of DDR-400, but I don't know if anyone has tested it yet.

There is also the possiblity that Apple may pair the fast bus with dual channel DDR-333, not ideal but more practical. They may even run the FSB slower than 900mhz. Who knows? Not us!

Edit: For all we know, Apple will use RDRAM. They could even use a custom nForce2... who knows.

 

[MacCoaster]

Originally posted by ddtlm
I'm a little confused what you are saying, but the PPC-970 does support two 32-bit channels as a "FSB" and each clocks at 900mhz. This is a little like how Rambus memory clocks at obscene rates, but only has 8-bit channels.

Read this original article I posted: http://arstechnica.com/wankerdesk/3q02/powerpc.html. Especially note this:

One of the most important and least-discussed features of the PowerPC 970 is its 900MHz DDR frontside bus. This bus physically runs at 450MHz, but it's double-pumped. Its architecture is interesting in that the bus actually consists of two, 32-bit unidirectional point-to-point links.