Advocacy
Dojo (HowTo)
Reference
Markets
Museum
News
Other
Myths
Press
General
Hack
Hardware
Interface
Software
Standards
People
Forensics
Web
CodeNames
Easter Eggs
History
Innovation
Sightings
Opinion
Martial Arts
ITIL
Thought
|
I keep getting lots of questions about articles over at MacObserver / Webintosh.
- http://www.macobserver.com/news/99/march/990312/powerpc.html
- http://www.macobserver.com/news/99/march/990319/powerpc2.html
This article will respond to the first of the articles.
Technical Analysis Of The Mac Processor Techology, Part I
Tuan Truong has some valid points in the articles -- but the truth gets obfuscated in the opinion side of the article. There is not much balance in the article -- not because of what Tuan says, but because of what he does not say. So I will break apart the articles and try to explain things more clearly.
To convert customers from a dominant platform to a smaller niche platform, a common adage is that the niche platform must be 2 to 5 times more powerful, cheaper, or more productive. "Twice the performance for half the [production] cost" of a comparable Intel Pentium was the rallying cry and the expectation of the PowerPC microprocessor partners: Apple, IBM and Motorola, during the heady days of it birth. Apple moved it's primary business to it based on that expectation.
Some valid points. In general you need to have at least double the performance to really get people to consider migrating -- on that we agree. Five times is a bit much.
At the start of the PowerPC alliance, AIM said that they had two goals -- one was to produce a comprable processor to what Intel had, but at half the price/power/size. The other goal was to have a high end processor that would be twice the performance of what Intel had. Both would be done at any given time. AIM has consistently met the first goal -- but the second goal has been spotty. They've probably achieved about 30 - 50% faster, in the real world -- but rarely hit the double goal. Now there are variants of the PowerPC that are achieving those goals (like the Power3) -- but they aren't exactly what AIM had in mind.
Part of the problem was that AIM announced exactly what speed and performance goals they were going to hit for the next 4 or 5 years -- and they did an amazingly good job of forecasting. But that announcement gave Intel the chance to target as well. The P6 core was rushed out the door, and basically was able to nullify the high end claims, and achieve something close to parity -- by hugely increasing complexity and cost (or processor).
Many people confused processor costs with system costs -- and that was misperception on their part. Processors are only one variable in a system (though a fairly large one).
That expectation did not come to fruition, and it seems as if that will still be the case in the future. As of now, the performance of the PowerPC and the Intel Pentium family of processors are in a virtual dead heat in performance and price. There are advantages and disadvantages for each platform, but none approaching the goal of being "twice the performance for half the cost". The following is a set of performance numbers I have gathered for the current PPC 750, the upcoming G4 PPC, and the Pentium II/III on their respective web sites. I'm using base numbers, not peak, as much as I can....
There are many "flaws of omission" in this paragraph.
First, SPEC is a really bad suite for measuing performance now days. For example it has very little ability to handle parallelism (either single thread or SIMD type operations). So in some ways Intels SSE and MMX are not getting factored in -- and the G4s far superior SIMD (AltiVec) is also not getting factored in. The suite is also not very good at measuring DSP/NSP (Signal Processing) type functions -- which PowerPC does far better than x86. So if you compare a photoshop filter there are many cases where a slower (MHz) PowerPC just stomps on the PentiumIII -- which contradicts the SPECmark theoretical numbers. There are many other reasons why SPEC has flaws. The truth is that BYTEmarks also measure performance and state that the PowerPC is over twice as fast -- mainly because a lot of the BYTE measurements are good at using NSP/DSP type functions, and use real world compilers, and other things that help the PowerPC, but in some ways are also reflective of the real world. So the truth is that SPEC seems to be biased against the PowerPCs real world advantages. Application tests, and other benchmarks (which include GLINT and many others) contradict SPECs value. PC people and many techie types love to quote SPEC, but I find it biased. Of course BYTE is not much better -- it to me is the equivelent of the same amount of BIAS, but this time in the PowerPCs favor. So for all practical matters I don't count BYTE as the complete truth either. In fact, I pretty much mentally average many benchmarks (especially Application benchmarks) -- and find that there is a performance difference of about 30 - 70% in the PowerPCs favor, and I just average that to about 50% better -- not as much as BYTE claims, and not the almost complete parity that SPEC does. This seems to be reflective of most real world Application tests with the exception of Microsoft Office -- but Office has a sort of Windows emulator (call remapper) that makes it a poorer performer on the Mac.
So as has become cliche -- "there are lies, damn lies, and statistics". People use the numbers that they want to -- and use those to support their position de jour. Others just use bad benchmarks because they don't know better. The reality is that the PowerPC is better, almost all agree on that, the only debate is how much better -- from 5 - 10% all the way up to 100% or more. As mentioned I believe it is about 50% better.
By the end of 1999 Intel and Motorola are expected to have moved the PowerPC G4 and Pentium III class of processors to a 0.18 micron process enabling them to clock their processors in the 600 MHz to 800 MHz range. IBM will probably stay with their brand new CMOS7S 0.20 micron process with copper interconnects in the near future, but they have yet to include their Silicon-on-Insulator technology which should enable them to clock their PPC processors in the 600 MHz to 800 MHz range as well.
The problem is I don't think Tuan is "plugged in" and he doesn't really know what is going on. Motorola is likely to move to .20 - .18 and copper more in the mid-year range (not end of 1999), and IBM is likely to add SOI to their chips in the same time frame. IBM is probably a few months behind Motorola for G4s -- but I'd expect to see fast G3s around mid year.
Since many real world results, like photoshop tests, show that SSE and MMX both only give the PentiumIII parity with the G3, any additions the G4 adds are likely to be significant. Ironically, for SPEC the G4 should give a near doubling of some FP (Floating Point) numbers -- that realistically most users will likely notice these gains much less (more support for the theory that SPEC has little real world value). AltiVec should offer serious performance boost for many things -- especially as there is likely to be more OS support for it, than their is for SSE and MMX on the PC.
NOTE: SSE and MMX seem to give a nice doubling of performance to Pentiums -- which enabled them to catch up to the signal processing performance of the PowerPC. But I've seen some results with a 14 TIMES improvement by adding AltiVec -- though I expects a more modest 2 - 4 times performance boost will be more common. With more OS and Application support, it is quite likely that those boosts will result in far more visible gains than what SSE and MMX offer (combined).
The floating point in the G4 is a healthy improvement to the PPC 750/604e and is clearly faster than the Pentium III. But the lack of improvement in integer performance is disappointing and indicates to me that the Motorola G4 is just an evolutionary enhancement to the PPC 750 plus the addition of AltiVec unit. Overall, the preliminary performance numbers do not come close to the "2 to 5 times better" goal for moving people off a dominant platform to a much smaller competing one. I expect the Pentium III to be comparable to the Motorola PPC G4 in the personal computer space. The PPC G4 should perform admirably in the embedded markets where the AltiVec will dominate.
Remember, SPEC is unlikely to show the performance advantages that the PowerPC has in integer -- and for some tasks (like say Photoshop, Sound, Video) there will be many cases where the G4 more than exceeds the 2 times better goal. The 5 times better goal is just a construct of the author, so I'll ignore that one. I don't expect the PentiumIII to be "comprable" to the G4 -- it is just that the G4 is not enough better to drag too many users out of their holes, and have them choose the G4. But just because some will not change does not mean that the processor (or machine) is not better.
If the G4 was the only "add" that Apple had in store for people this year, that would be the end of the story. But the author is either unaware of, or chooses to ignore, the many other changes in store for Macs this year. New OS flavors, like 8.6 (Veronica), 8.7 / 9.0 (Sonata) and OS X Client (late this year or early next), and the release of OS X Server, are all likely to give a big boost to the real world performance of PowerPCs. Probably 50 - 100% performance boosts. Apple is also prepared to reenter the MP (MultiProcessing) arena this year -- and likely to mainstream it and reduce costs of MP substantially with MultiCored processors (likely the G5) coming out early next year. And there are other hardware adds -- like changes to the memory subsystem which can increase performance (significantly) as well. So there is a good chance that by late this year, or early next, that the PowerPC system will easily enter the "2 - 5 times" faster arena that the author is talking about -- he just only had a partial picture of Apple's and Motorola's goals.
What seems worse is the sad state of the memory and I/O architectures supporting the PowerPC microprocessors. Neither IBM nor Motorola seems to produce an AGP/PCI bridge and memory controller chipset. The Yosemite architecture is fine, very good, except for the lack of AGP and being able to address 2 GB of RAM which are limitations in the current Motorola MPC106 "Grackle" PCI bridge/memory controller chipset. I haven't found much on what IBM and Motorola are planning for the PPC memory and IO architecture, only that the G4 will eventually move to a 128 bit system bus, while Intel is driving towards AGP 4x, Rambus, PC133 SDRAM, PCI-X, and NGIO. This may be indicative of Motorola's and IBM's focus on the embedded markets for their PPCs. The market does not need such flexible and high performance technologies. This is where Apple will have the most problems. If you can't feed the processor, you won't get the maximum performance out of the processor.
Interesting paragraph. It makes a lot of mistakes that many PC magazines make -- they start comparing the PCs NEXT generation stuff against THIS generation of Mac stuff. Yosemite uses the Grackle I/O chip, which doesn't offer AGP -- but the Grackle is still competitive with Intels BX chip set in the real world. Hmmmm. Show me the real world performance advantages of the PentiumIII and I'll listen -- but real world comparisons tend to show that all these theoretical advantages are not resulting in any serious gains.
Intels next generation chip set is like to support faster memory and later chip sets may support faster PCI busses. Nothing revolutionary there -- we know that technology marches on. But the Sawtooth machine and Uni-North bridge chip (as reported by Mac OS Rumors and other sources) are likely to include AGP and far faster memory access (and maybe technologie like PC133) -- probably BEFORE Intel ships their next generation chip set. So why does the author assume that the PowerPC won't improve on at least a parity basis? Not only that, but the PowerPC has a bigger L1 and L2 cache than the Pentiums (and faster access to those caches), so it is less memory starved than the Pentiums in the first place. Those points seem to be ignored.
NOTE: According to rumors the UniNorth is not likely to support 128 bit bus (too much added cost), but will still likely double (or more) memory performance through interleaved memory accesses (multiple requests on the bus at the same time). If you looked at the specs (MHz or width) of the BUS, you'd see no substantial change -- yet in the real world you will likely see a far larger change. Apple doesn't pre-announce this stuff because it can impact sales -- but Intel doesn't sell computers (just processors), so they are free to pre-announce what is coming -- and some will continue to compare those announcements to what others are shipping.
Lastly the author made a mistake of assuming that only IBM and Motorola are producing memory and I/O controllers. It seems that Apple is likely to produce their own (which is specialized for the tasks the Mac needs to do). IBM and Motorola are likely to focus on support chips that support their needs (embedded and some high end stuff). So the author takes his ignorance of the situation (about Apple's memory chips) and makes bad assumptions and conclusions.
Just like back in the late 80s and early 90s where Motorola dropped the ball in improving the 68k architecture, Motorola and IBM are both dropping the ball with PPC development. By concentrating on the embedded market rather than the personal computer market they are failing to improve it at a faster rate than the x86 line. If Apple's plans are to compete and gain market hare against the "Wintel" and Linux/x86 world, this lack of PPC and PPC subsystem development becomes a major pitfall in that plan. The PPC platform will have parity, but it will not have the generational leap in performance nor price advantage to make it easy to gain market share. With Mac OS 10 (Consumer), Apple has a very portable operating system given that Carbon was designed and implemented right. I think it would be time for Apple to consider moving away from PPC, Motorola at least, and look at another architecture or another strategy.
I don't disagree that Motorola was dropping the ball on the 68000 architecture. But that is decade old news. I also agree that IBM and Motorola have both failed to offer the level of support and drive that they should have with the PowerPC. Both managements have been pretty daft at points -- and in fact it is likely that Motorolas incompetant management (and lack of foresight) has been leading to many of the PowerPC architects leaving and "seeking opportunities elsewhere". So I think it is a little false for Motorola to blame Intel for all those problems -- if they had a clue, offered comprable salaries, and had some vision (and reasonable plans) then it would be far harder for Intel to seduce processor people away.
But the fact still are that despite some idiocy in both Motorola and IBM management, Intel is no better (and from the industry stories I've heard are actually worse). Despite the problems, the PowerPC has been out improving the x86 line -- and is likely to accellerate that. Apple is going to compound that advantage with serious improvements in the OS as well. And because of the lower power requirements, lower cost, higher performance, and more open market of the PowerPC (remember, the x86 is pretty proprietary and Intel does everything they can to make it more so), Apple has been gaining in mindshare and marketshare.
Saying that Apple should jump ship (at this point) is not a very wise idea. Maybe if IA64 (Merced or McKinley) deliver on their promises, Apple should look to that on the high end (and slow migration). If Apple and SUN had merged maybe combining SPARC and PowerPC would have been doable. But right now there is really no advantage to changing, and HUGE disadvantages. Apple has a lot of advantages by staying, like multi-core chips (because of the lower size and power requirements), Mac Portables that consistantly outperform (by the 2 - 5 times numbers the author desires) the PC / x86 counterparts, and far lower costs for Apple.
I keep hearing that Apple is getting it's processors for double digits. Which would mean that G3s and G4s are likely to cost Apple about 1/5 to 1/6th of what comprable Pentiums would cost. Of course Apple has to turn that money back into R&D because Motorola and IBM have dropped the ball in some areas -- but it is certainly a very large processor advantage, even if that hasn't made it into a huge system cost advantage -- yet.
|
