Slashdot Mirror
Preliminary OS X & PPC 970 Benchmarks
Dixie_Flatline writes "Macbidouille.com is reporting that they have
preliminary benchmarks involving PPC970 hardware. The results are seriously impressive. We're looking at a single processor PPC 970 1.4GHz machine quite strikingly outperforming a dual G4 1.42GHz machine. Don't worry, there's an English translation embedded in the page so you don't have to try to muddle through the French." Update: 05/05 19:58 GMT by T : Thanks to Eric from macbidouille.com, above link updated to a static page; hopefully you'll get better response this way.
Merci de votre patience et de votre compréhension.
,mémoire centrale. La carte mère Mach 64 est optimisée au maximum pour l'usage de la DDR-SDRAM.- Le PPC 970 ne perd en aucun cas du temps en exécutant des applications 32 Bits.
:
Vous allez comprendre en lisant ces tests pourquoi il nous était impossible de les publier avant. Maintenant que nous avons appris que les ventes de G4 pro sont anémiques, la publication de ces tests ne risque plus d'avoir d'incidence sur le marché. Cette publication ne fera plus qu'une chose, inciter les MacUsers qui passent au PC en désespoir de cause à attendre pour acheter un Mac.
[By reading these benchmarks you'll understand that we couldn't publish them before.
Now we know that PM G4 sells are stuck at a very low level, the following test results won't have much incidence. It will however make the ones switching to PC wait for the next generation of Power Macs.]
Ces premiers tests datent de mi Mars 2003. Ils ont été réalisés sur un modèle de présérie à 1,4 GHz. Le système était une Alpha de Panther en version 7B5 et 7B8 optimisée 64 Bits mais les applications testées étaient en 32 Bits.
[The first benchmarks were done during March 2003 on a preview model running at 1.4 GHz. OS was an alpha version 7B5 and 7B8 of Panther, optimised for 64 bits processor, but the applications tested were only using 32 bits.]
Sous Photoshop, le PPC 970 Mono 1,4 est 87% plus rapide qu'un Dual G4 1,42 GHz.
Sous Final Cut Pro, le PPC 970 Mono 1,4 est 112% plus rapide qu'un Dual G4 1,42 GHz.
Sous Alias|Wavefront Maya Render, le PPC 970 Mono 1,4 est 254% plus rapide qu'un Dual G4 1,42 GHz.
[Photoshop : PPC 970 mono 1.4 is 87% faster than a Dual 1.42 GHz Final Cut Pro : PPC 970 mono 1.4 is 112% faster than a Dual 1.42 GHz Alias|Wavefront Maya Render : PPC 970 mono 1.4 is 254% faster than a Dual 1.42 GHz]
Cette seconde série de tests a été réalisée sur des machines sorties de l'usine et donc identiques à celles qui seront en vente. Notez qu'il n'y a pas encore de certitude sur la mise en vente du modèle haut de gamme Dual 2.0 GHz, car la disponibilité en volume suffisants de ces puces n'est pas encore certain. Il reste donc possible qu'Apple ne fasse une gamme Mono 1,4,Dual 1,6, Dual 1,8 GHz.
[The second series of benchmarks were done on the same computers that will be sold. There is however a doubt on the presence of the up-market dual 2.0 GHz as the availability of these chips isn't sure. It seems Apple will surely be able to sell Mono 1.4 GHz, Dual 1.6 and Dual 1.8.]
Le commentaire est simple. Le PPC 970 relègue le G4 au rang de machines de secrétaire.
[The result is that the G4 compared to the PPC 970 is now a secretary computer.]
Voici les explications de ces résultats:
- L'altivec démontre une amélioration de performances de 80% sur le 970. Mais ce n'est pas à cause de la puce en elle même, mais grâce à l'accès extrèmement rapide du processeur à la
- L'optimisation de la carte mère est telle que le passage du mono au biprocesseur permet pratiquement de doubler la puissance effective. On arrive à 90% de performances en plus contre 50 pour le G4.
[A few explanations to the results
- The Altivec shows a 80% increase of performances with the 970. This is not due to the chip itself, but to the high speed access between processor and central memory. The Mach 64 motherboard is highly optimised for the use of DDR-SDRAM.
- There is no performance loss when the PPC 970 executes some 32 bits apps.
- The motherboard optimization almost allows dual processors to reach double performance. In fact it's about 90% efficiency gained with the second processor, compared to 50% for the G4.]
Lorsque l'on voit ces résultats on comprend mieux pourquoi
Mirror list
Mouse powered Chips, Open source Processors and Lego
Well, if you'd looked at the bar charts in the artcle, you'd have seen that the 1.4 GHz
benchmarks at about the same or a little faster than a 3 GHz P4.
No electrons were harmed creating this post, though some may have been subjected to electrical and/or magnetic fields.
It's in the article.
Test 1: Cinema 4D-XL
PPC G4 Dual 1.42 GHz 33 seconds
Pentium IV 3.0GHz 30 seconds
PPC 970 1.4 GHz 29 seconds
PPC 970 Dual 1.8 GHz 18 seconds
Test 2: Photoshop Actions
PPC G4 Dual 1.42 GHz 73 seconds
Pentium IV 3.0GHz 58 seconds
PPC 970 1.4 GHz 50 seconds
PPC 970 Dual 1.8 GHz 24 seconds
Test 3: Bryce 5
PPC G4 Dual 1.42 GHz 21 seconds
Pentium IV 3.0GHz 16 seconds
PPC 970 1.4 GHz 16 seconds
PPC 970 Dual 1.8 GHz 7 seconds
Actually, macbidouille.com is known for its ACCURACY on rumours. They had early photos of the Quicksilver PowerMac, they had photos of prototype motherboards for XServe, they were true about the specs of 2002 Apple-Expo Macs, etc...
Note that it is NOT a rumours site, but a Mac news / hacks site.
The story states explicitly that the OS and hardware are 64-bit, but the applications are 32-bit.
Ho hum. Another /.'ing. Here's a mirror of of the French, and one for the Babelfish-translated English.
Mirror
"Mes amis fans de Mac, notre attente va être récompensée. Nous avons fini de défendre une cause difficile. Apple va devenir le roi du monde !"
Actually its more like:
My friends, fans of the mac. Our wait will be rewarded. We will finish defending a difficult cause. Apple will become the king/ruler of the world.
They kinda left out a sentence. My French is a little weak these days..
Here
For those of you replying to this message.
This is a troll, the same comment pops up on all the Mac stories.
rationalists do it by the rules
empiricists do it to the rules
Oh, really? Somebody better tell the boys at IBM.
You need to learn some more about the PPC arcitecture.
PPC was designed from the ground up to scale to 64bit without affecting the performance of 32bit apps on the same processor. 32 and 64bit comingled apps can live quite happily on the same machine. There is no porting or special software required.
When a developer gets around to porting their app(s) to 64bitness, they can take advantage of newer features and higher performance.
The 32/64 bit conversion should go at least as smoothly as all the others: System 6->System 7+,68K->PPC,G3->G4, OS9->OS x. In each case the developer was under no pressure to release (properly written) software specially compiled for the new arcitecture, the hardware and/or OS masked the change and allowed the older apps to "just work".
Article X: The powers not delegated... by the Constitution...are reserved...to the people
IBMs vecotor processer is Altivec compatible. Apple calls it velocity engine. I think they just called it altivec since some people wouldn't know that Altivec is just Motorolla's name for their vector processor. When these things are in Macs, apple will probably call them G5s with velocity engine, even though they will be fourth generation(or second if you count how long they've been 64-bit) chip and not the exact same thing that was velocity engine.
All this seems very nice.
Lets get it out now.
The thing to remember is that the PowerPC is originally based on the IBM POWER chip -- a native 64 bit chip that can do 32 bit programs as well.
IBM tends to undersestimate and overproduce. They arent just making it for Apple, they will put the 970 in their own Linux blade servers and NetVista boxes for financial stuff. Also, the 970 is a variant of the POWER4 dualcore Risc monster processor in IBMs big server iron.
IBM doesnt screw around. Motorola is becoming irrelevant.
Heres another key reason why this chip might actually be as fast as MacBidoulle claims:
The system bus runs at 900 MHZ. The current mac system bus runs at 167 mhz. Think about it. A 900 lane highway vs a 167 lane highway? This chip will have monstrous bandwith. And the power consumption will be reduced a big deal as well..
Look at this official IBM presentation from last october
and this ArsTechnica review as well
The 970, being a 64 bit chip, allows more memory than 4GB, the current 32 bit limit. Servers need more than 4 gigs, especially IBM's monster iron.
10 years ago my Mac used 32 MB's of ram. Now its up to 768 megs. Sooner or later, it will go past 4 gigs. Better to get this transition done now than later.
The current PPCs (The g4s) are wide, but shallow. The much faster Pentium 4s are deep but narrow.
This is a guess, and if any cpu engineer wants to help out, id appreciate it.
The P4 stuffs all execution data down the pipe as fast as it can. If there's a break in the chain of execution instructions, the whole chain must be shoved down the pipe again.
The G4 spreads it all out over multiple pipes, but the pipes arent deep. The main work is figuring out which pipe is free to shove stuff into.
This is a gross simplification, so please bear with me.
The 970, on the other hand, has more pipes than the G4 and the Pentium 4, but the pipes are deeper than the P4. So it can stuff a whole ton of stuff down and be very efficient. Wide and deep. Theres a bit of a tradeoff, but the chip is just engineered much better.
I read the Ars technica article a long time ago and the IBM PDF file a while ago too. I would not be suprised if the data on Mac Bidoulle is accurate.
I am waiting for apple to stuff a 970 into a PowerBook, preferably the 15 inch one. I am waiting on that for my next computer. I do not want the G4. The Mobo on the G4 just doesnt have a wide enough bus to suck up massive amounts of data. The 970 mobo will.
The 970 mobo will be 900 Mhz. Intel has the 533 mhz mobo and soon will have the 800 mhz mobo.
Motorola and Apple were fighting about how to make the data path on the mobo. Motorola had the chips, they were just being strange. Motorola's problems stunted apple with the g4 for a long time. Apple had to overclock the g4 so much that the g4 tower got obscenely loud.
I welcome the 970 and want it in a Mac ASAP. I think that WWCD was delayed to show the developers the chip and a version of Panther that will have it. Bring it on! Lets see IBM take on Intel in the chipmaking business.
My bets are on IBM
It didn't sound that way to me, it's already common knowledge that the 3.0 P4 beats the 1.4 G4 at most stuff. The impression I got was that they didn't want to deflate Apple's Power Mac sales by pointing out that a much faster processor is due shortly. But then they found that G4 PMac sales are already in the tank, so it wouldn't do any harm (which is also common knowledge, so that doesn't entirely make sense). Of course this is assuming they're not fabricating numbers out of thin air.
How to solve most of our problems: 1.Lots of nuclear plants. 2.Cure aging.
The Panther builds are not on the B build train yet, so this article is wrong. There is no such thing as Panther 7B6.
I went and looked it up because I was not 100% sure.
Power point presentation
This power point presentation in pdf form shows on page 10 that yes there are seperate 32-bit modes. It also explains the 'basic' differences.
DeveloperWorks page
This article explains in fairly basic manner the difference between 32 and 64 bit assembly under linux and from this you can derive that there are 2 seperate modes for the PPC 970 64-bit implementation. Also if you go and look at the PPC 64-bit ABI document at the bottom of the page you'll realize that the 64-bit mode cannot run 32-bit code.
Note that all the length of all instructions, regardless of mode (32,64) is always 32-bit. So just because you're running in 64-bit mode doesn't mean that all your instructions are twice as long (contrary to popular opinion). Just that all your addresses are twice as long.
Anyway, I don't want to come off as a PITA. But it bugs me that people don't understand the difference between 64 and 32 bit computing and really there isn't much difference at all in many cases.
the web site is not a credible source and you would do well to take any "news" from it with a grain of salt
the 970 is a chip shrouded in mystery and there is no way a site like that is going to have access to any real benchmarks
that said, the 970 is not vapor and it will come, but when it will come and at what speeds and whether or not it will even be used in Macs is still all speculation
Maybe is because there is no dual Pentium 4 in the market. If you want dual processing, you need to go with Xeon.
Instruction size doesn't change.
Remember, when the PPC spec was set down over a decade ago, the 32 -> 64-bit transition was planned for. The PowerPC architecture is a 64 bit architecture with a 32 bit subset. All of the instructions are 32 bit, but some of them operate on 64 bit data. Really, there is no need for more than 2^32 or roughly 4 billion instructions. I don't know what the total instruction count for PPC is, but I'm sure it is less than 500. Altivec alone is 162 instructions
Anyway, the current G4 PPCs have 32-bit integers, 64-bit floating points, and 128-bit vectors. The 970 will have 64-bit integers, 64-bit floating points, and 128-bit vectors. The only change is the integer unit and the bus width. There are new instructions for operating on 64-bit integers, but that is it for new instructions. The 970 will be able to handle 32-bit integers with no problem.
--
The internet is the greatest source of biased information in the history of mankind.
I enjoy reading Hannibal's CPU articles on ars, but (if this isn't an impostor) what he's saying here doesn't make any sense:
Floating-point: the 970 will spank the G4e
Integer: The G4e will spank the 970
Let's look at the SPEC scores:
Disclaimer: The PPC 970 scores are IBM's stated estimates only... though IBM tends to under-estimate, if anything. Also, I could not find official G4e scores from Motorola, but the ones here were referenced in several places on CPU web sites.
SPECInt2000:
PPC 970 @ 1.8 GHz: 937
G4e @ 1.4GHz: 418
G4e @ 1.8 GHz (scaled): 537
Based on those numbers, I certainly wouldn't say the G4e spanks the 970 on integer performance! In fact, I'd say exactly the opposite. Hannibal rightly says that the 970 spanks the G4e on FP; those numbers are even more skewed!
SPECFP2000:
PPC 970 @ 1.8 GHz: 1051
G4e @ 1.4GHz: 248
G4e @ 1.8 GHz (scaled): 319
So it looks to me like the 970 almost doubles the G4e's performance on integer, and more than triples it on FP. I'll guess I'll have to wait for Hannibal's ars article to see why he reaches the conclusions he does.
P.S. Despite these numbers, I don't believe Macbidouille.com's posted application benchmarks are real. But I do believe Apple will use the 970 before the year ends, and I do believe that the 970 is going to be a huge improvement over the G4. Though I personally get along with my 667Mhz machine just fine...
You suck?
Seriously, I've got a Blue & White G3/500 and it doesn't take anywhere near 20 seconds for menus to drop down. Most of the time it is less than a second, most of the time it is too fast for me to time. If it takes that long for menus to drop down, you must be doing something seriously wrong.
Hmmmm... my B&W G3 400Mhz runs Jaguar 10.2.5 just dandy. Photoshop 7 is noticeably in OS X than Photoshop 6 was in Classic or booted into OS 9. Yeah, Office X is a bit of a dog, but that's alright. I pretty much just use it for Word and Entourage anyway.
Granted, when I use the DP 1Ghz G4 at the school where I teach Final Cut Pro with a nice new Nvidia GPU, I notice a huge speed difference over OS 9 and OS X on my lowly G3 at home. For everyday web graphics use though, Jaguar running PS 7 and Illustrator 10 is great on my B&W. And I've had absolutely NONE of these 20 second drop down window problems. EVER. I did back in the Public Beta, but I really hope you're not forcing that on your poor users.
Seriously, you're doing something wrong. 10.2 was like giving a new life to my B&W. 10.0 and 10.1 weren't even useable on a G3, but Jaguar, well... basically it rocks.
Pooty tweet
Here's a fixed link (with HTML in it): PowerPC 970 Annoucement
Ok, I see that I should've clarified a bit more, but I was in a hurry to get out of the house so I just fired off a quick post. First off, see this thread:
http://arstechnica.infopop.net/OpenTopic/page?q= Y& a=tpc&s=50009562&f=8300945231&m=3470943335&p=5 6
Check out pages 52 onward for some detailed discussion of these issues in advance of my article.
Now for a preliminary explanation (let's see if I can condense this):
In a very small nutshell, the 970 has less general-purpose integer hardware than either the G4e or the P4. It has two general-purpose ALUs (or arithmetic logic units, which do integer computation) that are both mostly symetric. This means that both ALUs handle almost all types of integer ops with a two cycle latency. However, there are some differences, but more on that in a sec.
The G4e, on the other hand, has one complex integer unit and three simple integer units. The three simple integer units have a one-cycle latency and handle all the basic types of integer instructions (add, multiply, etc.). Longer, more complex multi-cycle instructions, of which there are few and these show up statistically more rarely than the fast integer ops, are handled in the complex ALU.
So a basic comparison of ALU hardware shows you that the G4e has slightly more integer hardware that's more specialized and hence potentially faster. (Think a supermarket with two general purpose checkout lanes vs. a supermarket with three express lanes and one general purpose checkout lane).
This doesn't tell you the whole story. First, the good: The 970 handles CR logical operations in a separate unit, the CR logical unit. These types of ops are done on the G4e in the complex integer unit. So this bit of specialization helps the 970 out just a bit, but only a bit because CR ops are relatively rare.
Now for the bad, which is a killer: the 970's group dispatching scheme dictates that one ALU is fed from dispatch slots 0 and 3, while the other is fed from dispatch slots 1 and 2. (If you don't know what a dispatch slot is, reread my first 970 article.) So of the four possible integer ops that can be dispatched in parallel on any given cycle to the 970's ALU issue queues, two are constrained to go to one unit and two are constrained to go to the other. This sort of partitioning scheme makes code scheduling critical, because if there's a mix of integer ops and other types of ops (e.g. loads, stores, etc.) then one ALU's issue queue(s) could be oversubscribed while the others' languishes, due to the fact that the other ops happen to be pushing all the integer ops into one particular pair of dispatch slots (i.e. either 0 and 3 or 1 and 2).
Now, this is potentially bad enough already. But when you factor in the fact that the ALUs are not symetrical, and that certain types of ops can only go to one ALU and hence MUST go into one of only two dispatch slots, then you get a recipe for further choking of dispatch bandwidth.
Ok, I've probably managed to confuse anyone who's read this far, but so be it. You asked for an explanation. Read that thread I linked above for more discussion, or just wait for my article (it should be finished any day now) for a more user-friendly explanation with nice color diagrams and such.
The end result is that the 970's ALU hardware is weaker than that of the G4e, the P4, and the Athlon. So its clock-for-clock integer performance will be worse, at least this is what I'm predicting. We'll see if I'm right.
Now, this really isn't too big a deal to my mind, because most people care more about floating point and vector ops for the types of desktop and workstation apps that run on a Mac.
More worrisome is the inferior Altivec (or, as IBM calls it, VMX) hardware. The G4e has a superior and more robust SIMD implementation, but it's severely hobbled by a lack of FSB and memory subsystem bandwidth. I'm sure that IBM will improve the SIMD situation in future releases of the chip, though. Right n
Senior CPU Editor | Ars Technica | http://arstechnica.com/
An acronym is simply a word consisting of the initial letters of the words of some phrase.
An abbreviation is a shortening of a single word, not multiple words.
The two examples you give (LASER, SCUBA) were originally, printed in all caps, even though they were pronounced as single words right from the start.
Acronyms are often printed in all caps to make it clear that they are acronyms. Only many years of usage, and the consequent common knowledge that they are acronyms, results in their being printed in lower case.
Comparing what the power4 gets on specint 2000 @ 1.45Ghz (Score of 935) versus Intel's P4 @ 3.06Ghz (Score of 1091), the power4 holds its own in integer performance and in fact, so should the 970. (AMD's athlon 3000+ comes in at 995, btw.) To claim that "clock-for-clock integer performance will be worse" is utterly bogus. (What are you talking about with resepect to a lack of FSB performance? The 970's 900Mhz blows away intel...)
Give the microarchitectures (and performance numbers) a glance next time before going on a dilettantish ramble about architectures.