What is needed is to deliver the perfect balance of integer to FPU to SIMD performance, and maybe Intel knows this.
No such thing. There are as many "perfect" balances as there are uses for these machines.
Quote:
Maybe Apple can see that the Intel processors are better balanced.
Apple's transition is based on an evaluation of future Intel processors vs. future IBM processors. This will have included all aspects of machine performance, including power consumption and cost(s). The better integer performance is certainly a big plus, and the inferior FPU/VPU performance is a minus (unless Intel's roadmap demonstrates those are better in the future too). Intel claims they are making FPU improvements to the Pentium-M chip line (which Apple will adopt).
Have to violently disagree with you. If an app has 4 threads, a dual 970mp PowerMac using the current G5 motherboard will destroy a Yonah desktop in performance, all of the 970mp cores each have 1 MB on-die L2 cache afterall.
I think we're starting to talk past each other-- I don't care all that much about the capabilities of the processors' execution units. Today's single-processor Pentium M 2.1GHz chips have 2MB of L2 cache per chip, and Intel has shown a great deal of expertise in doing good cache implementations both on-chip and on motherboard. Today's single-processor G5 has 512KB of L2 cache, and Apple has a history of comparably bad motherboard caching and chipsets. If I were a betting man, I'd put my money on Intel to create an architecture that can feed more cores with data for their next generation of chips. Today's CPUs tend to be slightly starved as it is... adding 2 more cores without a dramatic redesign is going to simply offer no benefits.
As for the next generation-- Yonah vs. 970mp-- and how they compare, it will be interesting. Intel is pulling more desktop technology into Yonah, and shoring up some of the Pentium M's weak points, while at the same time lowering power consumption and voltage. The G5 will be boosted to 1MB of cache per core, while the Yonah will have 2MB of cache shared between the cores, which will probably offer a significant win for the Yonah without VERY careful thread affinity algorithms on the G5. (In the worst-case scenario of two threads working on the same data, the G5 has effectively half the L2 cache as the Yonah and twice the memory bus overhead.)
In summary, I suspect the real limitation of multi-core performance will be memory busses and caching, where the Yonah appears to be ahead of the 970mp. I don't think anyone can predict whether a 2-core Yonah will beat a 2-core 970mp, a dual 2-core 970mp, or neither. I don't think any statements there are "ludicrous", although obviously someone's going to be wrong.
Multithreading - I would expect the Yonah to win. Intel now has way more experience with desktop dual-core than anyone else, except perhaps AMD. The quad-core G5 is a nice thought, but it would very much depend on the cache architecture to overcome the anemic RAM configuration, and Apple has shown no evidence to be clueful in this regard. (a single threaded app, in contrast, will not take advantage of multiple CPUs, so if you are just running that one single threaded app, then whether you have 1, 2, or 4 cores is irrelevant)
Vector operations - I would expect the G5 to win for scientifically-targeted vector operations, and the Yonah to win for games-oriented vector operations. A lot of opportunities to vectorize games that exist in SSE3 aren't as easy to use (ie. slower) with Altivec with all the data massaging required. On the other hand, Altivec really does scream for certain imaging and scientific apps.
Floating point - G5 definitely wins from what we know so far. It's a super chip for floating point.
Memory bandwidth - Yonah, hands down. Intel is 667MHz DDR2 now, while Apple is stuck at 400MHz DDR. Bidirectional pipes to the CPU are nifty and all, but to then bottleneck them with slow RAM is silly.
Disk access - I don't see how either CPU has an advantage here. Disk speeds will be the bottleneck for virtually any bus or CPU. However, PCI-E is sure as heck better than Apple's northbridge/southbridge at shuffling the data between motherboard components here.
Graphics Card acceleration - Yonah win, easily. AGPx8 is a slow, slow bus compared to PCI-E.
Multithreading: You appear to have misunderstood my point. I am not talking about Hyperthreading, multicores, or anything hardware related. I am simply saying that a multithreaded app will take advantage of multiple CPUs (or multiple cores). Therefore, a well threaded app running on 4 G5 cores clocked at 2.5 Ghz or higher will easily outrun the same app running on 2 Pentium-M cores clocked around 2 Ghz (the targeted clock speed for Yonah's release).
While it may be technically true, it is also disingenuous to state the Intel has more dual core experience than anyone else. What is that, two or three months of experience? IBM has been making dual core chips since they released Power4 in 2001 - they are the pioneers of dual core technology!
Finally, if you think by "quad-core G5" I mean a 4 cores on a single die, that is wrong. A G5 machine with four cores will have two 970MP processors, with two cores per processor. A dual-core Powerbook will have one dual-core Yonah.
Vector operations - Games oriented vector operations, blah blah blah. Get real, pretty much any game that seriously challenges a 2+ Ghz G5 or Pentium-M is going to also require a good graphics card. We don't know exactly what GPU technology will be available next year, but if you look at what's out there today, the fastest chipset that could realistically fit in a Powerbook is the Mobility X700 (the Mobility X800 and the Geforce Go 6800 are only designed for 10+ lb "desktop replacement" laptops...they suck way too much power for a Powerbook). The Mobility X700 isn't bad, but it will be completely slaughtered, chewed up, and spit out by high end desktop cards like the Radeon X850 or GeForce 7800. If you think a 5 lb laptop GPU is ever going to come close to competing with a high end desktop GPU, you are smoking some pretty strong stuff.
Memory bandwidth - Yonah's FSB is expected to be 667 Mhz (167 Mhz quad-pumped). EACH 970 MP will have a 1250 Mhz E-Bus (625 Mhz double-pumped), so even when only one core is being used AND all of the memory usage is in only one direction, the G5 system will still have similar bandwidth to Yonah (5.2 GB/sec for Yonah, 4.9 GB/sec for the G5). Of course, that is the worst case scenario. In general usage, the G5 system will have up to FOUR TIMES the FSB bandwidth of the Yonah system (9.8 GB/sec per 970MP times two 970MPs).
Furthermore, most laptop chipsets do not use dual channel memory. There is no guarantee the Powerbook will.
As for DDR vs DDR2, you are assuming that the G5 chipset will not support DDR2 by next year. You are also ignoring the latency issues with DDR2 that make DDR2/533 benchmark worse than DDR400 in most cases.
Disk access - Hint: it's not in the CPU. Remember, the original poster was claiming that a 2006 PowerMac 970MP is going to be slower than a 2006 Yonah Powerbook. This is demonstrably untrue for many, many important applications.
The PowerMac can use a 4 channel RAID, or even XServe RAID, for hundreds of MB/sec of bandwidth. The Powerbook will at best have a single, small 7200 rpm mobile HD which might max out at 40 or 50 MB/sec at most. Again, the Yonah Powerbook gets slaughtered. And PCI-E is irrelevant...the limiting factor is clearly the drive (and after that, the drive interface), not 133 Mhz PCI vs. PCI-E.
Graphics Card acceleration - Okay, now you're just being completely nuts. You really do believe a POS Mobility X700 is going to outrun an AGP X850 or AGP GeForce 6800GT just because the former has PCI-E. WRONG. The X850 and 6800GT are going to ABSOLUTELY DESTROY the Mobility chip. To think otherwise is to have absolutely no understanding of how GPUs work and what AGP/PCI-E actually do. (let me guess, you also think the PCI-E Radeon X300 will beat an AGP Radeon X850)
Clearly, all of my original points stand. I will assume that you are not completely ignorant, but that you simply had no idea what systems were actually being compared. Next time READ THE THREAD before you respond to someone's post. It will save you some embarrassment!
It's almost like saying a Yonah laptop will be faster than the dual 2.7 GHz PowerMac G5 Apple currently ships. That's pretty crazy when we know Yonah will only be at 2.2 GHz. [/B]
Hmm, that's a good point. Except he is trying to compare a 4 core G5 to a 2 core Powerbook, so what he's really saying is that a single core 2 Ghz Dothan lightweight laptop should be outrunning a two core Dual G5/2.5
And yet strangely, if we assume that a 2 Ghz Dothan is about equivalent to a 3 Ghz P4 (actually, the Dothan is probably a bit slower for the class of applications we're talking about, which benefit from vector, floating point, and Hyperthreading enhancements on the P4, but it won't be a huge difference) we can see that the G5 would be:
2.4 times faster in After Effects rendering
1.9 times faster in Cinema 4D rendering
1.9 times faster in Photoshop MP actions
1.4 times faster in Photoshop SP actions
2.1 times faster in Bryce (actually, I think the Dothan would do somewhat better than the P4 here, but it's not going to change anything in the end)
In other words, the Dual G5 wins in every single benchmark, often by 2x or more.
The Dothan would also lose in games. Obviously there is no Dothan laptop gaming equivalent on Barefeats, but you can compare the Dual G5 FPS numbers to Dothan/Mobility X700 numbers from other sites
Unreal 2K4: G5/X800 gets 60/137 FPS (flyby/botmatch), Dothan/X700 gets 20 FPS (31 FPS with no AA). (this one from PC Mag)
Note that these numbers seriously overstate the Dothan/X700 performance, because the G5 is always tested at 1600x1200 (4x AA where applicable), whereas the Dothans are tested at 1400x1050 for Halo and 1280x1024 for Doom and Unreal.
I don't think anyone can predict whether a 2-core Yonah will beat a 2-core 970mp, a dual 2-core 970mp, or neither.
Watch me. I will predict that the fastest 4 core 970MP G5 system shipping in summer 2006 (assuming Apple releases such a beast) with the best Mac ATI or NVidia card will beat the fastest Yonah-based Powerbook that Apple ships by summer 2006 (assuming that Apple sticks to its Intel Mac shipping schedule), in a majority of the following applications:
Photoshop (say, an average of the Barefeats MP and SP filter tests)
Cinebench
After Effects
Doom3 at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
Halo at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
and a whole bunch more, but these ones we can be pretty sure that we will see benchmarks from at Barefeats.
If you are so sure that it is very hard to predict whether 970 MP PowerMac will beat Yonah Powerbook, then you should be happy to take this bet at odds better than 50/50. In fact, I will offer you THREE-TO-ONE odds in your favor, so this should be a great bet for you. I'd be gladly willing to wager $30,000 on this (i.e., if you win you get $30,000, if I win I get $10,000). How much would you be willing to wager?
Watch me. I will predict that the fastest 4 core 970MP G5 system shipping in summer 2006 (assuming Apple releases such a beast) with the best Mac ATI or NVidia card will beat the fastest Yonah-based Powerbook that Apple ships by summer 2006 (assuming that Apple sticks to its Intel Mac shipping schedule), in a majority of the following applications:
Photoshop (say, an average of the Barefeats MP and SP filter tests)
Cinebench
After Effects
Doom3 at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
Halo at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
and a whole bunch more, but these ones we can be pretty sure that we will see benchmarks from at Barefeats.
If you are so sure that it is very hard to predict whether 970 MP PowerMac will beat Yonah Powerbook, then you should be happy to take this bet at odds better than 50/50. In fact, I will offer you THREE-TO-ONE odds in your favor, so this should be a great bet for you. I'd be gladly willing to wager $30,000 on this (i.e., if you win you get $30,000, if I win I get $10,000). How much would you be willing to wager?
What if the 970 MPs never get released? Do you automatically lose?
I would like to purpose an observation that could be supported by the dev kits. We have for years been saying that the Mac is tuned for everyday life stuff but not for bench testing. Is it possible that we could look at the Intel processors as being somewhat the same in that they may be well balanced for desk top processing.
I'm fairly confident that the development kits are faster or seem faster primarily for 2 things. The user interface performance (app startup, window creation, filesystem) is driven by integer performance and burst memory performance, and the Pentium 4 660 in the dev kits is very good at that. Intel in general does very well in burst memory performance.
If they start benching Photoshop, FCP and whatnot, a 2.7 GHz 970fx should be about the same as the 3.6 GHz P4 660. If there was a GUI speed benchmark, I think the P4 660 would win hands down. A 2 GHz Dothan should win hands down against a 2, maybe 2.2 GHz 970fx.
Photoshop (say, an average of the Barefeats MP and SP filter tests)
Cinebench
After Effects
Doom3 at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
Halo at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
I suspect the G5 will win in Photoshop. I'm not sure why you limit the Doom3 and Halo to 1280x1024 or above... do you think the Yonah PowerBook would win with lower resolutions? In any case, I think the Yonah will be competitive on the games, even considering laptop GPU suckage. Don't know much about Cinebench or After Effects characteristics, so can't say.
And what about browsing, checking Mail, using iPhoto, copying data over a network, opening/closing/resizing windows, doing searches, or any of a hundred other things the typical user will want to do with their laptop?
I also have no clue what sort of chipset Apple will go with. From history, they could easily hobble their PowerBooks simply to maintain market segmentation. In short, I think the technology will be comparable, but I'd say the odds are worse than 3-to-1 that Apple will screw it up again and squander the opportunity. You also specify the best ATI or nVidia graphics card available... which Apple never ships as a factory configuration anymore.
So I'm not going to take your bet, because 1. I'm not the betting type, and 2. as I maintained earlier, I have no idea whatsoever how these things will benchmark against each other and I don't want to lose money over such a silly argument.
If there was a GUI speed benchmark, I think the P4 660 would win hands down.
I'm curious about the real profiles here -- the slow part of the MacOS X GUI has always (apparently) been Quartz 2D. Quartz 2D is a pixel rasterizer built around a floating point display model. This means it is mostly likely going to be primarily a combination of FPU and vector work, and moving a lot of memory. The Pentium4 will do pretty well here because this stuff pipelines nicely to take full advantage of clock rate, and the P4's fast bus and caches will be a big benefit. The (existing) Pentium-M will look a weak by comparison. Depending on the balance between memory usage and FPU/VPU compuation the G5 should be strong, and the 970MP with its bigger caches ought to do noticably better. Without real profiling data it isn't possible to know...
I'm fairly confident that the development kits are faster or seem faster primarily for 2 things. The user interface performance (app startup, window creation, filesystem) is driven by integer performance and burst memory performance, and the Pentium 4 660 in the dev kits is very good at that. Intel in general does very well in burst memory performance.
If they start benching Photoshop, FCP and whatnot, a 2.7 GHz 970fx should be about the same as the 3.6 GHz P4 660. If there was a GUI speed benchmark, I think the P4 660 would win hands down. A 2 GHz Dothan should win hands down against a 2, maybe 2.2 GHz 970fx.
I know that a test at Tom's Hardware showes that Yohna can be faster than the fastest Itanium, and Pentiums. So I think that a Dual Yohna machine will do just fine against a 970MP. In that you just stated that a 2.7 970fx is comparable to a P4 660, or were you talking about a single processor? What I have seen, the Dual 970s run neck and neck with the single Pentiums and Itaniums. I don't think that the current Pentiums and Itaniums were designed with dual configurations in mind but that will change as Intel switches to Yohna and beyond.
I'm fairly confident that the development kits are faster or seem faster primarily for 2 things. The user interface performance (app startup, window creation, filesystem) is driven by integer performance and burst memory performance, and the Pentium 4 660 in the dev kits is very good at that. Intel in general does very well in burst memory performance.
If they start benching Photoshop, FCP and whatnot, a 2.7 GHz 970fx should be about the same as the 3.6 GHz P4 660. If there was a GUI speed benchmark, I think the P4 660 would win hands down. A 2 GHz Dothan should win hands down against a 2, maybe 2.2 GHz 970fx.
Since this 'old-tech' Dothan easily clocks upto 2.5 Ghz as shown here http://www.tomshardware.com/cpu/2005...ntium4-10.html you will note for all REAL benchmarks (rar,ripping,etc) it wins hands down over current gen P4 and AMD chips. So i believe in most daily comp chores this lil champ would be great. If the next gen deliver's with an uprated SIMD and possible future on die mem controller, we can stay cool in both thermal and performance.
I'm not at all understanding the statement "most computing tasks that are bandwidth limited are unidrectional". I can't imagine bandwidth sensitive applications where you simply suck vast amounts of data IN to a processor and have little to no output. I'd welcome education on that.
That's easy: its not true.
<snipping the rest which is not at issue.>[/B][/QUOTE]
I would strenuously say it actually is true. What are you looking at right now? That display has been created by many times more code and data read into the processor over the FSB than ever goes out. Any type of quality encoding uses many times the amount of input data to the encoded output. Your forte, gaming code, will ship 10's of times more code upstream directly to the graphics card over the bus due to textures. In comparison to the the bytes shipped upstream there, the poly info shipped downstream off the CPU and over to the graphics card pales in comparison. From there almost everything goes out to the monitor or local storage (VRAM), not back to primary storage (RAM). Searches? Tons of data comes in--does it match? No, throw it away unchanged. Yes, keep that one little slice of everything you looked through. I could go on for a long time here.
The POWER series and 970 style balanced elastic bus is superior for server type transactions where you expect to ship large amounts of data off the CPU, but not much comparison or computation, and can use the clear one way downstream bus to avoid clobbering the upstream bus when shipping those packets out. In an aside though, those outgoing packets only go part-way down that bus to RAM, they get diverted at the DMA module to the appropriate I/O interface, usually one of the networks. But in computation intensive applications you typically consume far more data (and instructions which only ever go upstream) than finished data you ship back to RAM or other downstream devices.
Since this 'old-tech' Dothan easily clocks upto 2.5 Ghz as shown here http://www.tomshardware.com/cpu/2005...ntium4-10.html you will note for all REAL benchmarks (rar,ripping,etc) it wins hands down over current gen P4 and AMD chips. So i believe in most daily comp chores this lil champ would be great. If the next gen deliver's with an uprated SIMD and possible future on die mem controller, we can stay cool in both thermal and performance.
So a Yohna chip is basically two Dothans on one die, with improvements. So a MB design that uses two Yohnas is a system that could be faster than a MB utilizing 4 Pentium 660s.
I'm curious if OSX for Intel supports hyperthreading?
OS X is multithreaded and supports multi-threaded applications as it has done so from the beginning. Hyper-threading (SMT) will just make this support more valuable.
OS X is multithreaded and supports multi-threaded applications as it has done so from the beginning. Hyper-threading (SMT) will just make this support more valuable.
Hyperthreading (within single core) is actually closer to SMP than SMT, but the scheduling is different than both.
I was wondering if OSX supports single hyperthreading. Maybe this is a non-issue with Yonnah.
I know that a test at Tom's Hardware showes that Yohna can be faster than the fastest Itanium, and Pentiums.
Lets stop including Itaniums as part of the discussion. In fact, lets just stop including Itanium in any discussion involving Apple all together. It's not proper until they enter the enterprise server market.
Quote:
So I think that a Dual Yohna machine will do just fine against a 970MP.
That wasn't the point of contention. Remember the original comment: a Yonah laptop will outrun a quad-PPC PowerMac, and will be the first Mac laptop in years to be faster than desktops. This essentially means a 2.2 GHz Yonah laptop versus a dual 2.5 GHz 970mp. It's an apples and oranges comparison at best. For things that Apple would ship a quad-PPC PowerMac for, it will crush a Yonah laptop at those things, and most of those things would involve multithreaded apps.
One of the things that I think we all agree on is that Yonah will perform single threaded spaghetti integer code better.
Quote:
In that you just stated that a 2.7 970fx is comparable to a P4 660, or were you talking about a single processor?
Single processor, assuming properly tuned code, and "comparable" meaning that for some things one processor will do much better than the other, but on the whole, roughly comparable.
The dual 2.7 970fx PowerMac is a competitive machine against dual x86 machines at its price point for Apple's market.
Since this 'old-tech' Dothan easily clocks upto 2.5 Ghz as shown here http://www.tomshardware.com/cpu/2005...ntium4-10.html you will note for all REAL benchmarks (rar,ripping,etc) it wins hands down over current gen P4 and AMD chips.
It's not surprising, but I will also note that a ~2.5 GHz Dothan and Yonah doesn't really exist. Dothan is limited to ~2.1 GHz, Yonah limited to ~2.2 GHz.
If Intel ships a ~2.5 GHz Sossaman, with appropriate core logic chipset support for desktop machines, then sure, Apple should use them. I'm sure they would rather switch earlier than later.
Quote:
So i believe in most daily comp chores this lil champ would be great.
Who said that it wouldn't?
Quote:
If the next gen deliver's with an uprated SIMD and possible future on die mem controller, we can stay cool in both thermal and performance.
Merom-based Conroe, the 8th generation IA-32 architecture (w/EM64T), looks to be when Apple will start switching the Power Macs to the Intel platform. We're all hoping it will be great. Heck, Sossaman could be good enough.
So a Yohna chip is basically two Dothans on one die, with improvements. So a MB design that uses two Yohnas is a system that could be faster than a MB utilizing 4 Pentium 660s.
Just a tiny nuance we should remember about Yonah and Dothan. Yonah is a basically two Dothan "cores" on a die with 2 MB of shared L2, as far as we know. Dothan has 1 core with 2 MB of L2. Text illustration:
Code:
Dothan:
----------------------
| --------- |
|| Dothan | |
|| core | |
| --------- |
| -------------------- |
|| ||
|| 2 MB L2 cache ||
|| ||
| -------------------- |
----------------------
Yonah:
----------------------
| --------- --------- |
|| Dothan || Dothan ||
|| core+ || core+ ||
| --------- --------- |
| -------------------- |
|| ||
|| 2 MB L2 cache ||
|| ||
| -------------------- |
----------------------
Whether a dual Yonah system is faster than a quad P4 660 system is all dependent on the clock rate of the CPUs. If Yonah is limited to 2.2 GHz, probably not. Competitive in some things, but on the whole, probably not.
Back to the Yonah vs 970mp comparison. Consider the improvement from the 970fx to the 970mp: 2x the cache per processor. The 970fx has 512k cache, but each core of the 970mp has 1 MB of dedicated L2. Grand total of 2 MB on the die. Twice the cache is the equivalent to about adding 5 to 10% in clock rate (within limitations), so each core of the 970mp could be about as fast a 2.7 GHz 970fx. That's pretty good. If Apple used dual channel DDR2-667 for a dual 2.5 GHz 970mp machine, that would be some good performance improvements for streaming memory apps.
Yonah compared to Dothan on the other hand, is more complicated. Single-threaded performance will be like Dothan, perhaps a little better due to the "digital media boost" improvements Intel is doing. But multithreaded performance, hmm, the cores will be a little bit slower because they have to share the same cache. A Yonah core is almost be like a Dothan with only 1 MB L2 and slightly slower to the real 2 MB L2 Dothan. So, it's not exactly like 2 Dothans on one die.
The multicore speedup factor for Yonah compared to Dothan won't be or shouldn't be as good to the multicore speedup factor for a 970mp compared to the 970fx.
Comments
Originally posted by Apple
Time to move on...
*sniff, sniff* Are you breaking up with me? - IBM
Originally posted by 00100011
*sniff, sniff* Are you breaking up with me? - IBM
Apple to IBM "I need some space. Just wanted to tell you it's not you but me"
Originally posted by Brendon
What is needed is to deliver the perfect balance of integer to FPU to SIMD performance, and maybe Intel knows this.
No such thing. There are as many "perfect" balances as there are uses for these machines.
Maybe Apple can see that the Intel processors are better balanced.
Apple's transition is based on an evaluation of future Intel processors vs. future IBM processors. This will have included all aspects of machine performance, including power consumption and cost(s). The better integer performance is certainly a big plus, and the inferior FPU/VPU performance is a minus (unless Intel's roadmap demonstrates those are better in the future too). Intel claims they are making FPU improvements to the Pentium-M chip line (which Apple will adopt).
Originally posted by THT
Have to violently disagree with you. If an app has 4 threads, a dual 970mp PowerMac using the current G5 motherboard will destroy a Yonah desktop in performance, all of the 970mp cores each have 1 MB on-die L2 cache afterall.
I think we're starting to talk past each other-- I don't care all that much about the capabilities of the processors' execution units. Today's single-processor Pentium M 2.1GHz chips have 2MB of L2 cache per chip, and Intel has shown a great deal of expertise in doing good cache implementations both on-chip and on motherboard. Today's single-processor G5 has 512KB of L2 cache, and Apple has a history of comparably bad motherboard caching and chipsets. If I were a betting man, I'd put my money on Intel to create an architecture that can feed more cores with data for their next generation of chips. Today's CPUs tend to be slightly starved as it is... adding 2 more cores without a dramatic redesign is going to simply offer no benefits.
As for the next generation-- Yonah vs. 970mp-- and how they compare, it will be interesting. Intel is pulling more desktop technology into Yonah, and shoring up some of the Pentium M's weak points, while at the same time lowering power consumption and voltage. The G5 will be boosted to 1MB of cache per core, while the Yonah will have 2MB of cache shared between the cores, which will probably offer a significant win for the Yonah without VERY careful thread affinity algorithms on the G5. (In the worst-case scenario of two threads working on the same data, the G5 has effectively half the L2 cache as the Yonah and twice the memory bus overhead.)
In summary, I suspect the real limitation of multi-core performance will be memory busses and caching, where the Yonah appears to be ahead of the 970mp. I don't think anyone can predict whether a 2-core Yonah will beat a 2-core 970mp, a dual 2-core 970mp, or neither. I don't think any statements there are "ludicrous", although obviously someone's going to be wrong.
Description of Yonah cache:
http://www.tomshardware.com/hardnews...02_143758.html
Originally posted by Booga
Multithreading - I would expect the Yonah to win. Intel now has way more experience with desktop dual-core than anyone else, except perhaps AMD. The quad-core G5 is a nice thought, but it would very much depend on the cache architecture to overcome the anemic RAM configuration, and Apple has shown no evidence to be clueful in this regard. (a single threaded app, in contrast, will not take advantage of multiple CPUs, so if you are just running that one single threaded app, then whether you have 1, 2, or 4 cores is irrelevant)
Vector operations - I would expect the G5 to win for scientifically-targeted vector operations, and the Yonah to win for games-oriented vector operations. A lot of opportunities to vectorize games that exist in SSE3 aren't as easy to use (ie. slower) with Altivec with all the data massaging required. On the other hand, Altivec really does scream for certain imaging and scientific apps.
Floating point - G5 definitely wins from what we know so far. It's a super chip for floating point.
Memory bandwidth - Yonah, hands down. Intel is 667MHz DDR2 now, while Apple is stuck at 400MHz DDR. Bidirectional pipes to the CPU are nifty and all, but to then bottleneck them with slow RAM is silly.
Disk access - I don't see how either CPU has an advantage here. Disk speeds will be the bottleneck for virtually any bus or CPU. However, PCI-E is sure as heck better than Apple's northbridge/southbridge at shuffling the data between motherboard components here.
Graphics Card acceleration - Yonah win, easily. AGPx8 is a slow, slow bus compared to PCI-E.
Multithreading: You appear to have misunderstood my point. I am not talking about Hyperthreading, multicores, or anything hardware related. I am simply saying that a multithreaded app will take advantage of multiple CPUs (or multiple cores). Therefore, a well threaded app running on 4 G5 cores clocked at 2.5 Ghz or higher will easily outrun the same app running on 2 Pentium-M cores clocked around 2 Ghz (the targeted clock speed for Yonah's release).
While it may be technically true, it is also disingenuous to state the Intel has more dual core experience than anyone else. What is that, two or three months of experience? IBM has been making dual core chips since they released Power4 in 2001 - they are the pioneers of dual core technology!
Finally, if you think by "quad-core G5" I mean a 4 cores on a single die, that is wrong. A G5 machine with four cores will have two 970MP processors, with two cores per processor. A dual-core Powerbook will have one dual-core Yonah.
Vector operations - Games oriented vector operations, blah blah blah. Get real, pretty much any game that seriously challenges a 2+ Ghz G5 or Pentium-M is going to also require a good graphics card. We don't know exactly what GPU technology will be available next year, but if you look at what's out there today, the fastest chipset that could realistically fit in a Powerbook is the Mobility X700 (the Mobility X800 and the Geforce Go 6800 are only designed for 10+ lb "desktop replacement" laptops...they suck way too much power for a Powerbook). The Mobility X700 isn't bad, but it will be completely slaughtered, chewed up, and spit out by high end desktop cards like the Radeon X850 or GeForce 7800. If you think a 5 lb laptop GPU is ever going to come close to competing with a high end desktop GPU, you are smoking some pretty strong stuff.
Memory bandwidth - Yonah's FSB is expected to be 667 Mhz (167 Mhz quad-pumped). EACH 970 MP will have a 1250 Mhz E-Bus (625 Mhz double-pumped), so even when only one core is being used AND all of the memory usage is in only one direction, the G5 system will still have similar bandwidth to Yonah (5.2 GB/sec for Yonah, 4.9 GB/sec for the G5). Of course, that is the worst case scenario. In general usage, the G5 system will have up to FOUR TIMES the FSB bandwidth of the Yonah system (9.8 GB/sec per 970MP times two 970MPs).
Furthermore, most laptop chipsets do not use dual channel memory. There is no guarantee the Powerbook will.
As for DDR vs DDR2, you are assuming that the G5 chipset will not support DDR2 by next year. You are also ignoring the latency issues with DDR2 that make DDR2/533 benchmark worse than DDR400 in most cases.
Disk access - Hint: it's not in the CPU. Remember, the original poster was claiming that a 2006 PowerMac 970MP is going to be slower than a 2006 Yonah Powerbook. This is demonstrably untrue for many, many important applications.
The PowerMac can use a 4 channel RAID, or even XServe RAID, for hundreds of MB/sec of bandwidth. The Powerbook will at best have a single, small 7200 rpm mobile HD which might max out at 40 or 50 MB/sec at most. Again, the Yonah Powerbook gets slaughtered. And PCI-E is irrelevant...the limiting factor is clearly the drive (and after that, the drive interface), not 133 Mhz PCI vs. PCI-E.
Graphics Card acceleration - Okay, now you're just being completely nuts. You really do believe a POS Mobility X700 is going to outrun an AGP X850 or AGP GeForce 6800GT just because the former has PCI-E. WRONG. The X850 and 6800GT are going to ABSOLUTELY DESTROY the Mobility chip. To think otherwise is to have absolutely no understanding of how GPUs work and what AGP/PCI-E actually do. (let me guess, you also think the PCI-E Radeon X300 will beat an AGP Radeon X850)
Clearly, all of my original points stand. I will assume that you are not completely ignorant, but that you simply had no idea what systems were actually being compared. Next time READ THE THREAD before you respond to someone's post. It will save you some embarrassment!
Originally posted by THT
It's almost like saying a Yonah laptop will be faster than the dual 2.7 GHz PowerMac G5 Apple currently ships. That's pretty crazy when we know Yonah will only be at 2.2 GHz. [/B]
Hmm, that's a good point. Except he is trying to compare a 4 core G5 to a 2 core Powerbook, so what he's really saying is that a single core 2 Ghz Dothan lightweight laptop should be outrunning a two core Dual G5/2.5
And yet strangely, if we assume that a 2 Ghz Dothan is about equivalent to a 3 Ghz P4 (actually, the Dothan is probably a bit slower for the class of applications we're talking about, which benefit from vector, floating point, and Hyperthreading enhancements on the P4, but it won't be a huge difference) we can see that the G5 would be:
2.4 times faster in After Effects rendering
1.9 times faster in Cinema 4D rendering
1.9 times faster in Photoshop MP actions
1.4 times faster in Photoshop SP actions
2.1 times faster in Bryce (actually, I think the Dothan would do somewhat better than the P4 here, but it's not going to change anything in the end)
Source is http://www.barefeats.com/macvpc.html
In other words, the Dual G5 wins in every single benchmark, often by 2x or more.
The Dothan would also lose in games. Obviously there is no Dothan laptop gaming equivalent on Barefeats, but you can compare the Dual G5 FPS numbers to Dothan/Mobility X700 numbers from other sites
(for example, http://www.hothardware.com/viewartic...eid=637&cid=10 )
Halo: G5/X800 gets 57 FPS, Dothan/X700 gets 30 FPS
Doom3: G5/X800 gets 20 FPS, Dothan/X700 gets 16 FPS.
Unreal 2K4: G5/X800 gets 60/137 FPS (flyby/botmatch), Dothan/X700 gets 20 FPS (31 FPS with no AA). (this one from PC Mag)
Note that these numbers seriously overstate the Dothan/X700 performance, because the G5 is always tested at 1600x1200 (4x AA where applicable), whereas the Dothans are tested at 1400x1050 for Halo and 1280x1024 for Doom and Unreal.
Originally posted by Booga
I don't think anyone can predict whether a 2-core Yonah will beat a 2-core 970mp, a dual 2-core 970mp, or neither.
Watch me. I will predict that the fastest 4 core 970MP G5 system shipping in summer 2006 (assuming Apple releases such a beast) with the best Mac ATI or NVidia card will beat the fastest Yonah-based Powerbook that Apple ships by summer 2006 (assuming that Apple sticks to its Intel Mac shipping schedule), in a majority of the following applications:
Photoshop (say, an average of the Barefeats MP and SP filter tests)
Cinebench
After Effects
Doom3 at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
Halo at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
and a whole bunch more, but these ones we can be pretty sure that we will see benchmarks from at Barefeats.
If you are so sure that it is very hard to predict whether 970 MP PowerMac will beat Yonah Powerbook, then you should be happy to take this bet at odds better than 50/50. In fact, I will offer you THREE-TO-ONE odds in your favor, so this should be a great bet for you. I'd be gladly willing to wager $30,000 on this (i.e., if you win you get $30,000, if I win I get $10,000). How much would you be willing to wager?
Originally posted by bigmig
Watch me. I will predict that the fastest 4 core 970MP G5 system shipping in summer 2006 (assuming Apple releases such a beast) with the best Mac ATI or NVidia card will beat the fastest Yonah-based Powerbook that Apple ships by summer 2006 (assuming that Apple sticks to its Intel Mac shipping schedule), in a majority of the following applications:
Photoshop (say, an average of the Barefeats MP and SP filter tests)
Cinebench
After Effects
Doom3 at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
Halo at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
and a whole bunch more, but these ones we can be pretty sure that we will see benchmarks from at Barefeats.
If you are so sure that it is very hard to predict whether 970 MP PowerMac will beat Yonah Powerbook, then you should be happy to take this bet at odds better than 50/50. In fact, I will offer you THREE-TO-ONE odds in your favor, so this should be a great bet for you. I'd be gladly willing to wager $30,000 on this (i.e., if you win you get $30,000, if I win I get $10,000). How much would you be willing to wager?
What if the 970 MPs never get released? Do you automatically lose?
Originally posted by Brendon
I would like to purpose an observation that could be supported by the dev kits. We have for years been saying that the Mac is tuned for everyday life stuff but not for bench testing. Is it possible that we could look at the Intel processors as being somewhat the same in that they may be well balanced for desk top processing.
I'm fairly confident that the development kits are faster or seem faster primarily for 2 things. The user interface performance (app startup, window creation, filesystem) is driven by integer performance and burst memory performance, and the Pentium 4 660 in the dev kits is very good at that. Intel in general does very well in burst memory performance.
If they start benching Photoshop, FCP and whatnot, a 2.7 GHz 970fx should be about the same as the 3.6 GHz P4 660. If there was a GUI speed benchmark, I think the P4 660 would win hands down. A 2 GHz Dothan should win hands down against a 2, maybe 2.2 GHz 970fx.
Originally posted by bigmig
Photoshop (say, an average of the Barefeats MP and SP filter tests)
Cinebench
After Effects
Doom3 at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
Halo at any resolution of 1280x1024 or above, AA or no AA (both machines running same resolution/settings)
I suspect the G5 will win in Photoshop. I'm not sure why you limit the Doom3 and Halo to 1280x1024 or above... do you think the Yonah PowerBook would win with lower resolutions? In any case, I think the Yonah will be competitive on the games, even considering laptop GPU suckage. Don't know much about Cinebench or After Effects characteristics, so can't say.
And what about browsing, checking Mail, using iPhoto, copying data over a network, opening/closing/resizing windows, doing searches, or any of a hundred other things the typical user will want to do with their laptop?
I also have no clue what sort of chipset Apple will go with. From history, they could easily hobble their PowerBooks simply to maintain market segmentation. In short, I think the technology will be comparable, but I'd say the odds are worse than 3-to-1 that Apple will screw it up again and squander the opportunity. You also specify the best ATI or nVidia graphics card available... which Apple never ships as a factory configuration anymore.
So I'm not going to take your bet, because 1. I'm not the betting type, and 2. as I maintained earlier, I have no idea whatsoever how these things will benchmark against each other and I don't want to lose money over such a silly argument.
Originally posted by THT
If there was a GUI speed benchmark, I think the P4 660 would win hands down.
I'm curious about the real profiles here -- the slow part of the MacOS X GUI has always (apparently) been Quartz 2D. Quartz 2D is a pixel rasterizer built around a floating point display model. This means it is mostly likely going to be primarily a combination of FPU and vector work, and moving a lot of memory. The Pentium4 will do pretty well here because this stuff pipelines nicely to take full advantage of clock rate, and the P4's fast bus and caches will be a big benefit. The (existing) Pentium-M will look a weak by comparison. Depending on the balance between memory usage and FPU/VPU compuation the G5 should be strong, and the 970MP with its bigger caches ought to do noticably better. Without real profiling data it isn't possible to know...
Originally posted by THT
I'm fairly confident that the development kits are faster or seem faster primarily for 2 things. The user interface performance (app startup, window creation, filesystem) is driven by integer performance and burst memory performance, and the Pentium 4 660 in the dev kits is very good at that. Intel in general does very well in burst memory performance.
If they start benching Photoshop, FCP and whatnot, a 2.7 GHz 970fx should be about the same as the 3.6 GHz P4 660. If there was a GUI speed benchmark, I think the P4 660 would win hands down. A 2 GHz Dothan should win hands down against a 2, maybe 2.2 GHz 970fx.
I know that a test at Tom's Hardware showes that Yohna can be faster than the fastest Itanium, and Pentiums. So I think that a Dual Yohna machine will do just fine against a 970MP. In that you just stated that a 2.7 970fx is comparable to a P4 660, or were you talking about a single processor? What I have seen, the Dual 970s run neck and neck with the single Pentiums and Itaniums. I don't think that the current Pentiums and Itaniums were designed with dual configurations in mind but that will change as Intel switches to Yohna and beyond.
Originally posted by THT
I'm fairly confident that the development kits are faster or seem faster primarily for 2 things. The user interface performance (app startup, window creation, filesystem) is driven by integer performance and burst memory performance, and the Pentium 4 660 in the dev kits is very good at that. Intel in general does very well in burst memory performance.
If they start benching Photoshop, FCP and whatnot, a 2.7 GHz 970fx should be about the same as the 3.6 GHz P4 660. If there was a GUI speed benchmark, I think the P4 660 would win hands down. A 2 GHz Dothan should win hands down against a 2, maybe 2.2 GHz 970fx.
Since this 'old-tech' Dothan easily clocks upto 2.5 Ghz as shown here http://www.tomshardware.com/cpu/2005...ntium4-10.html you will note for all REAL benchmarks (rar,ripping,etc) it wins hands down over current gen P4 and AMD chips. So i believe in most daily comp chores this lil champ would be great. If the next gen deliver's with an uprated SIMD and possible future on die mem controller, we can stay cool in both thermal and performance.
Originally posted by Programmer
I'm not at all understanding the statement "most computing tasks that are bandwidth limited are unidrectional". I can't imagine bandwidth sensitive applications where you simply suck vast amounts of data IN to a processor and have little to no output. I'd welcome education on that.
That's easy: its not true.
<snipping the rest which is not at issue.>[/B][/QUOTE]
I would strenuously say it actually is true. What are you looking at right now? That display has been created by many times more code and data read into the processor over the FSB than ever goes out. Any type of quality encoding uses many times the amount of input data to the encoded output. Your forte, gaming code, will ship 10's of times more code upstream directly to the graphics card over the bus due to textures. In comparison to the the bytes shipped upstream there, the poly info shipped downstream off the CPU and over to the graphics card pales in comparison. From there almost everything goes out to the monitor or local storage (VRAM), not back to primary storage (RAM). Searches? Tons of data comes in--does it match? No, throw it away unchanged. Yes, keep that one little slice of everything you looked through. I could go on for a long time here.
The POWER series and 970 style balanced elastic bus is superior for server type transactions where you expect to ship large amounts of data off the CPU, but not much comparison or computation, and can use the clear one way downstream bus to avoid clobbering the upstream bus when shipping those packets out. In an aside though, those outgoing packets only go part-way down that bus to RAM, they get diverted at the DMA module to the appropriate I/O interface, usually one of the networks. But in computation intensive applications you typically consume far more data (and instructions which only ever go upstream) than finished data you ship back to RAM or other downstream devices.
Originally posted by ZoranS
Since this 'old-tech' Dothan easily clocks upto 2.5 Ghz as shown here http://www.tomshardware.com/cpu/2005...ntium4-10.html you will note for all REAL benchmarks (rar,ripping,etc) it wins hands down over current gen P4 and AMD chips. So i believe in most daily comp chores this lil champ would be great. If the next gen deliver's with an uprated SIMD and possible future on die mem controller, we can stay cool in both thermal and performance.
So a Yohna chip is basically two Dothans on one die, with improvements. So a MB design that uses two Yohnas is a system that could be faster than a MB utilizing 4 Pentium 660s.
Originally posted by skatman
I'm curious if OSX for Intel supports hyperthreading?
OS X is multithreaded and supports multi-threaded applications as it has done so from the beginning. Hyper-threading (SMT) will just make this support more valuable.
OS X is multithreaded and supports multi-threaded applications as it has done so from the beginning. Hyper-threading (SMT) will just make this support more valuable.
Hyperthreading (within single core) is actually closer to SMP than SMT, but the scheduling is different than both.
I was wondering if OSX supports single hyperthreading. Maybe this is a non-issue with Yonnah.
Originally posted by Brendon
I know that a test at Tom's Hardware showes that Yohna can be faster than the fastest Itanium, and Pentiums.
Lets stop including Itaniums as part of the discussion. In fact, lets just stop including Itanium in any discussion involving Apple all together. It's not proper until they enter the enterprise server market.
So I think that a Dual Yohna machine will do just fine against a 970MP.
That wasn't the point of contention. Remember the original comment: a Yonah laptop will outrun a quad-PPC PowerMac, and will be the first Mac laptop in years to be faster than desktops. This essentially means a 2.2 GHz Yonah laptop versus a dual 2.5 GHz 970mp. It's an apples and oranges comparison at best. For things that Apple would ship a quad-PPC PowerMac for, it will crush a Yonah laptop at those things, and most of those things would involve multithreaded apps.
One of the things that I think we all agree on is that Yonah will perform single threaded spaghetti integer code better.
In that you just stated that a 2.7 970fx is comparable to a P4 660, or were you talking about a single processor?
Single processor, assuming properly tuned code, and "comparable" meaning that for some things one processor will do much better than the other, but on the whole, roughly comparable.
The dual 2.7 970fx PowerMac is a competitive machine against dual x86 machines at its price point for Apple's market.
Originally posted by ZoranS
Since this 'old-tech' Dothan easily clocks upto 2.5 Ghz as shown here http://www.tomshardware.com/cpu/2005...ntium4-10.html you will note for all REAL benchmarks (rar,ripping,etc) it wins hands down over current gen P4 and AMD chips.
It's not surprising, but I will also note that a ~2.5 GHz Dothan and Yonah doesn't really exist. Dothan is limited to ~2.1 GHz, Yonah limited to ~2.2 GHz.
If Intel ships a ~2.5 GHz Sossaman, with appropriate core logic chipset support for desktop machines, then sure, Apple should use them. I'm sure they would rather switch earlier than later.
So i believe in most daily comp chores this lil champ would be great.
Who said that it wouldn't?
If the next gen deliver's with an uprated SIMD and possible future on die mem controller, we can stay cool in both thermal and performance.
Merom-based Conroe, the 8th generation IA-32 architecture (w/EM64T), looks to be when Apple will start switching the Power Macs to the Intel platform. We're all hoping it will be great. Heck, Sossaman could be good enough.
Originally posted by Brendon
So a Yohna chip is basically two Dothans on one die, with improvements. So a MB design that uses two Yohnas is a system that could be faster than a MB utilizing 4 Pentium 660s.
Just a tiny nuance we should remember about Yonah and Dothan. Yonah is a basically two Dothan "cores" on a die with 2 MB of shared L2, as far as we know. Dothan has 1 core with 2 MB of L2. Text illustration:
Dothan:
----------------------
| --------- |
|| Dothan | |
|| core | |
| --------- |
| -------------------- |
|| ||
|| 2 MB L2 cache ||
|| ||
| -------------------- |
----------------------
Yonah:
----------------------
| --------- --------- |
|| Dothan || Dothan ||
|| core+ || core+ ||
| --------- --------- |
| -------------------- |
|| ||
|| 2 MB L2 cache ||
|| ||
| -------------------- |
----------------------
Whether a dual Yonah system is faster than a quad P4 660 system is all dependent on the clock rate of the CPUs. If Yonah is limited to 2.2 GHz, probably not. Competitive in some things, but on the whole, probably not.
Back to the Yonah vs 970mp comparison. Consider the improvement from the 970fx to the 970mp: 2x the cache per processor. The 970fx has 512k cache, but each core of the 970mp has 1 MB of dedicated L2. Grand total of 2 MB on the die. Twice the cache is the equivalent to about adding 5 to 10% in clock rate (within limitations), so each core of the 970mp could be about as fast a 2.7 GHz 970fx. That's pretty good. If Apple used dual channel DDR2-667 for a dual 2.5 GHz 970mp machine, that would be some good performance improvements for streaming memory apps.
Yonah compared to Dothan on the other hand, is more complicated. Single-threaded performance will be like Dothan, perhaps a little better due to the "digital media boost" improvements Intel is doing. But multithreaded performance, hmm, the cores will be a little bit slower because they have to share the same cache. A Yonah core is almost be like a Dothan with only 1 MB L2 and slightly slower to the real 2 MB L2 Dothan. So, it's not exactly like 2 Dothans on one die.
The multicore speedup factor for Yonah compared to Dothan won't be or shouldn't be as good to the multicore speedup factor for a 970mp compared to the 970fx.