IBM's 90nm PowerPC 970FX

[Hattig]

Originally posted by nighthawk
Of course, this means that the 970 has a L3 cache up to 8GB running at the "same" speed as the P4 Extreme 2MB cache.

I think you need to read up on these things a lot more! That is totally incorrect.

For a start, the L3 cache on the P4EE will be extremely low latency, especially compared to normal memory. It will also be very wide - 256-bits most likely. That will increase its bandwidth.

And then the P4 has an 800MTps bus at 64-bits, i.e., 6.4GB/s total bandwidth. That pairs nicely with dual channel PC3200 memory.

The 970 has a 1GHz (at 2GHz core speed) bus at 32-bits in each direction, so it has a little more bandwidth than the P4 in total. However the P4 may have better prefetching hardware or any of the myriad other things that also affect these things, so you cannot draw a total conclusion just from bandwidth alone.

Athlon64/Opteron is the most optimal solution, with on-die memory controllers. That means that it connects to the memory controller at core clock speed (2.2GHz). I/O to the system or other processors is 800MHz DDR (1600MTps) at 16-bits in each direction as well.

 

[Samurai980]

Well all interesting reading. Personally I hope that IBM will be able to put out a processor that blows the Wintel world out in both theoretical performance and real performance in applications used today.

So when do the Quad 10GHz 980s, w/ 80GB of RAM, 2TB HD, ATI R420 Vid Card, that is liquid cooled and comes at the price of $500 come out? I need a G5 bad.

 

[Snowy_River]

Originally posted by rog
Well this is disappointing news. 7 months later and only 10% faster. Uh oh.

Uh, where do you get the 10% from? Is that just from the 1.1GHz bus speed? But we already know that the processor speed can be more than just a 2x factor of the bus. Suppose it's a 3x of a 800MHz bus. Oops, that's only a 20% increase. So, what if it's 3x of a 900MHz bus? Well, shucks, that's only a 35% increase. Well, maybe it'll actually be a 3x of a 1.1GHz bus. Then that would be 3.3GHz processor. Would that be enough to get you excited?

Also, the 2GHz 970fx is still too much of a power hog to be used in a powerbook. While the 1.4 GHz is better, who wants such a slow, single processor machine in 2004? Remeber the G5 is only about 20% faster than a same clock speed G4.

Oh, so a 20% boost in performance is worthless. Hmm. And how does the difference in performance from the FSB figure in? Surely that would increase the performance of the PBs.

Looks like they'll have to move to a 65nm process before fast G5s get into powerbooks. Very sad for pro portable users.

Oh, give me a break. 65nm is far enough away that it certainly won't be the next step. There will be a G5 PB on the 90nm process, whether it's with the 970FX or another update to the G5 (perhaps the G5m?) remains to be seen.

 

[tortoise]

Originally posted by stingerman
Another factor affecting performance is bandwidth and the G5 architecture has greater bandwidth. So when you are dealing with larger data sets, latency is a much smaller component of the performance equation.

No. Bandwidth is only more important than latency if you have a small number of very large objects in memory. If you have a larger number of smaller objects, the bandwidth is the same but latency dominates real throughput. For multimedia apps this is fine, as their memory usage tends to be dominated by large block transfers. For many high performance computing applications though, latency is the predominant scaling bottleneck.

Originally posted by stingerman
So to simply speak of latency as the driver behind performance is misleading. I believe that the G5 has the best balance of Latency, Bandwidth and bus frequency for an overall faster system. Couple that with IBM's advanced look ahead algorithms and it isn't an issue at all.

Most real-world supercomputing benchmarks (and I am not talking about the nigh useless LINPACK here) show that the mediocre memory latency of the PPC970 is its Achille's Heel. It has lots of bandwidth, but the latency is too high to let you use it all effectively. You say it isn't an issue at all, but impressively low latency is why the Opteron outperforms the G5 for more generic high-performance computing at the same clock speed. Clearly this will be an issue for some people.

If you are using your G5 for multimedia and such, it will be as fast as anything out there. Hell, it is almost perfectly tuned for that application. If you are working on very large datasets that have loads of non-sequential memory accesses (and many application spaces do) or more general high performance computing applications, it will be a less stellar performer. Since the focus of Apple tends to be on multimedia, the G5 works out very well for them. If you need high performance for other purposes though, it may not be the best choice.

 

[SiliconAddict]

OK. People are looking the XServer 2Ghz CPU to get a feel for what might be in a PowerBook. What I have to ask is what additional optimizations could IBM do to the G5 to make it more laptop friendly. The afore mentioned Pentium M shuts down entire sections of the CPU when not in use. I have to question if IBM and Apple may be working on such a thing for the G5. Lets face facts people. Two hours is pathetic for a laptop. I have to imagine Apple doesn’t want to fall anywhere under that time frame. (IMHO 2.5 hours should be minimum.) So lets say initial test batches of the 90nm chips were being produced late last fall/winter. Would it be conceivable that the end of year timeframe for a G5 PowerBook that Jobs alluded to last fall could be because of such optimizations?
Unfortunately my knowledge on the architecture of the G5 could fit on the head of an atom so I don’t know squat as to how much effort it would take to reengineer the G5 for notebooks. Is a year not enough time?

 

[acj]

No one has mentioned that the junction temp range is significantly higher. That is great for notebook uses. It means it can run stable at higher temps, requireing smaller heatsinks, fans, etc.

 

[stingerman]

Originally posted by tortoise
No. Bandwidth is only more important than latency if you have a small number of very large objects in memory.

Most real-world supercomputing benchmarks (and I am not talking about the nigh useless LINPACK here) show that the mediocre memory latency of the PPC970 is its Achille's Heel. It has lots of bandwidth, but the latency is too high to let you use it all effectively. You say it isn't an issue at all, but impressively low latency is why the Opteron outperforms the G5 for more generic high-performance computing at the same clock speed. Clearly this will be an issue for some people.

If you are using your G5 for multimedia and such, it will be as fast as anything out there. Hell, it is almost perfectly tuned for that application.

Hello, anyone home? You are just regurgitating my posts now and adding more troll crap.

In the real world the Big Mac Supercluster is the fastest supercomputer among all other off the shelf clusters. Including your superfast latency crap. What a troll... You have no idea how foolish your posts read.

 

[stingerman]

Originally posted by SiliconAddict
Two hours is pathetic for a laptop.

Consumer Reports had the latest PowerBook G4 at the longest lasting battery life of 4.5 hours. The 970FX is looking like it will use even less power.

As far as the 970FX PowerTune, it is looking like it has the following features:

1- Dynamic frequency based on usage and temperature (This is more efficient than Stepping.)

2- Dynamic Voltage based on current frequency. This provides a 4Times decrease in power usage and heat generation in addition to the linear decrease as a result of the frequency decrease.

3- Dynamically sleeping regions of the processor that are not in use.

4- Voltage Islands that allow for sections of the processor to use a lower voltage than other sections. (A technique born out of IBM's SOC work.)

5- eFuse, which allows IBM to switch between redundant portions of their processor in case a critical section fails. (Increases yield and life of the processor.)

 

[idkew]

Originally posted by isgoed
Hi Mr. American,

The correct spelling is "voilá". It is French.

signed,

Mr. European

I thought that was a musical instrument?

 

[pjkelnhofer]

What happened to the 750VX?

Amongst, all the dicussion of the 970FX, I wondered if anyone else noticed that the IBM Reference Guide contained no mention of the 750VX Chip. Which it seemed like a lot of people in here expected to be the chip in the next generation of iBooks and eMacs.

Anyone know what happened to it?

 

[Snowy_River]

Originally posted by SiliconAddict
...could fit on the head of an atom...

Uh, atoms don't have heads...

Originally posted by idkew
I thought that was a musical instrument?

And, that's a viola...

 

[Snowy_River]

Re: What happened to the 750VX?

Originally posted by pjkelnhofer
Amongst, all the dicussion of the 970FX, I wondered if anyone else noticed that the IBM Reference Guide contained no mention of the 750VX Chip. Which it seemed like a lot of people in here expected to be the chip in the next generation of iBooks and eMacs.

Anyone know what happened to it?

I don't know what happened to it, but I might point out that IBM only put out an announcement about the 970FX after the first Apple product using it was released (the Xserve). So, the 750VX may still be out there, just not being spoken of until it is released in the iMac/eMac/iBook line.

 

[stingerman]

Re: What happened to the 750VX?

Originally posted by pjkelnhofer
Amongst, all the dicussion of the 970FX, I wondered if anyone else noticed that the IBM Reference Guide contained no mention of the 750VX Chip. Which it seemed like a lot of people in here expected to be the chip in the next generation of iBooks and eMacs.

Anyone know what happened to it?

Steve Jobs stated that the G5 was the future of the Mac at the SF MacWorld. I think it would be smart to just use G5's across the board. This will greatly simplify optimizing the code for each platform and will allow Apple to get much better scaling and purchasing power. The diferentiator for the various lines seems to be more whether it is dual or single, how expandable is it, other components, etc.

I hope they go for G5 across the board.

 

[idkew]

Originally posted by Snowy_River
And, that's a viola...

i hope you realize that was a sarcastic comment.

 

[Snowy_River]

Originally posted by idkew
i hope you realize that was a sarcastic comment.

Of course I did. I was just having some fun!

 

[Rincewind42]

Re: About power consumption

Originally posted by manu chao
- 970FX, 90nm, 1.0v, typical, 1.4 Ghz
--> 12.3 Watt

Actually the 1.4 Ghz part appears to be running at 1.1v. If you do the math, extrapolating the 1.4Ghz part at 1v gives you a number about 20% too high (if you extrapolate 2 Ghz @ 1.3v) or low (if you extrapolate 1.4 Ghz @ 1v). As to what frequency you get when running at 1v, there doesn't appear to be any indication from the document that started this thread, but likely that is related to the speed ramping that the chip does (and which the original 970 could do too - that's how the current PMG5 does it's speed scaling).

 

[Rincewind42]

Re: What happened to the 750VX?

Originally posted by pjkelnhofer
Amongst, all the dicussion of the 970FX, I wondered if anyone else noticed that the IBM Reference Guide contained no mention of the 750VX Chip. Which it seemed like a lot of people in here expected to be the chip in the next generation of iBooks and eMacs.

Anyone know what happened to it?

It was a rumor . IBM never announced such a beast, and if you do a search on IBM's site, you will (oddly, as I didn't expect this) find only one link that leads to an AppleInsider article on it. Look for yourself.

Personally I always thought that the 750VX was a silly rumor - why spread your efforts on two products when you can get more out of those efforts on one product?

 

[jettredmont]

Re: About power consumption

Originally posted by manu chao
Note that Intel uses the term 'Thermal design', and I have no idea whether that corresponds to typical or max, or something in between.

"Thermal Design" is the typical power consumption/heat dissipation figure, not the Max. The notes on that PDF state this.

 

[Chryx]

Originally posted by nighthawk
Originally posted by csubear

Also, the data cache is only 32kb, the instruction cache uses a separate 64kb. The spec sheet for the P4 Extreme does not list a separate amount for instruction cache. I believe that this is one of the weaknesses of the x86 architecture -- it does not include a separate cache for instruction data.

Nope, the P4 has a 'trace cache' when it stores post CISC>RISC translated code, this takes the place of an L1 instruction cache.

it's around 21KB, but Intel haven't released an exact number.

And 'no seperate caches for instructions' is most definately NOT an "x86 weakness", the Athlon (XP/64/FX etc) have 64KB to instructions and 64KB to data at the L1 level.

What you aren't accounting for, is that caches are as much about the actual latency of the cache as they are the the size, the P4 needs low latency caches to keep the pipeline fed, so it has very small, very fast caching, the Athlon and the 970 both have less extreme pipelines and don't need the same extreme latency, so they have more cache.

As for why they have the amounts they have, both AMD and IBM will have run typical code mix simulations to work out the best balance of cache size, cache latency and die area required.

 

[Chryx]

Originally posted by tortoise
<snip>


Would you care to explain why the G5 is so devastatingly quick for scientific tasks like fluid dynamics then?

 

[pjkelnhofer]

Re: Re: What happened to the 750VX?

Originally posted by Rincewind42
It was a rumor . IBM never announced such a beast, and if you do a search on IBM's site, you will (oddly, as I didn't expect this) find only one link that leads to an AppleInsider article on it. Look for yourself.

Personally I always thought that the 750VX was a silly rumor - why spread your efforts on two products when you can get more out of those efforts on one product?

Thanks, I guess I never saw the original thread, I just kept seeing the "750VX" talked about in iBook forums. I had assumed it was a real chip. Maybe it will be, but not any time soon, I would guess.

 

[stingerman]

Originally posted by Chryx
Originally posted by tortoise
<snip>


Would you care to explain why the G5 is so devastatingly quick for scientific tasks like fluid dynamics then?

He's a troll or just completely out of his league.

 

[stingerman]

Re: Re: Re: What happened to the 750VX?

Originally posted by pjkelnhofer
Thanks, I guess I never saw the original thread, I just kept seeing the "750VX" talked about in iBook forums. I had assumed it was a real chip. Maybe it will be, but not any time soon, I would guess.

I can see Apple developing a Mac on a chip by taking advantage of of IBM's advanced SOC design tools. They can take a 750 ALU core, add on a 970 type FP2 unit, a VMX Altivex unit, then put in the memory controller, the device controllers, the GPU and communications parts all on one chip. It would allow Apple to make for a 1/2" notebook and maybe a full blown OS X handheld.

The important thing about the Apple/IBM relationship is that Apple's designers have a lot more tools at their disposal and a lot less limitations. It will be interesting what comes out of their R&D using IBM's capabilities.

 

[Chryx]

Originally posted by stingerman
The P-M is actually a modified old P-III.

I don't think it is, Intel have been keeping fairly quiet about its design, but it doesn't seem to be directly based on the P3 from what I can gather.

it more closely resembles the P3 than it does the P4, but then, the Athlon more closely resembles the P3 rather than the P4 as well.


Whatever it is though, IMO it's the best design to come out of Intel in a VERY long time.

 

[jettredmont]

Originally posted by rog
Oh yes, I'm such a troll for not jumping for joy and drooling over every extra MHz that IBM/Apple can give us. Not everybody on this forum blindly praises Apple for everything they do. Sorry, but I'm telling it like it is. While the G5 is a drastic improvement over the G4, it hasn't gotten any faster in 7 months, it's not in the portable or consumer lines, and it isn't contributing to growth of the platform. It also hasn't resulted in price cuts needed to grow market share. The news posted at the top of this thread is simply underwhelming. In the real world (please join it) Apple is competing with PCs, not all of which run on celerons and windoze, that are drastically cheaper and faster, and the vast majority of consumers barely know what an OS is or what the implications of one are, or why OSX is vastly superior to anything else. Get over yourself.

1) There was no mention of speed increases here. The change is a change in process size from 130nm to 90nm.

2) Apple does not (ever) continually release upgraded spec machines. It is not akin to the Wintel world in this respect. There will be no speed bump until the day that speed bump is announce. As such, seven months with 0% speed increase is normal and expected. Wait until the speed bump. Then, if it's really only 10% (to 2.2GHz) then you can complain.

3) Apple has said it will achieve 3GHz shipping by the end of this summer. Steve Jobs doesn't like being wrong. I doubt he would make such a claim were he not 90% positive that it would happen. That's a 50% increase in one year, which is quite respectable.

4) Many more consumers know the difference between OSX and Windows than know the difference between a Celeron and a G4. Granted, they may not know *all* the differences, but they sure as hell can point out that the OSX screen looks just a little bit different from the Windows XP Teletubbies Desktop!