Anybody know when 12GB DIMMs will be available? 12x8!
I'm really wanting to put 96GB of RAM in my Mac Pro. (That would be the one I haven't ordered yet because I don't have the money.) 
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Nehalem Mac Pros Arrive: Unboxing and Benchmarks
- Thread starter MacRumors
- Start date
- Sort by reaction score
Anybody know when 12GB DIMMs will be available? 12x8!
I'm really wanting to put 96GB of RAM in my Mac Pro. (That would be the one I haven't ordered yet because I don't have the money.) 
Please read this doctoral dissertation from Cambridge University on the challenges of SMT scheduling, and tell me again that very little needs to be done.
Be sure, it will never be the case that multiple cores scale with 100% per core. That's simply not possible. Grand Central will just reduce dependencies in OS routines and therefore it might improve the performance of some applications by 2-5%, but sure, not more. It will help speed up some single core applications to gain also some % from multiple cores through the usage of system routines.
However, even if an application is able to scale the work 100% with each task, the number of cores will reduce the scaling performance. Therefore REAL 16 cores might be only speed up good scaled software by ~35-50% compared to REAL 8 cores if nothing else of the architecture is better. You will lose a lot with interprocess communcication, e.g. to ensure that the cache lines keep udated in the right way.
Sure, it always depends on your usage and I know that some people really need every percentage of performance. Those should always vote for the fastest processor with the maximum available cores. However, if you have to decide between a faster Mac Quad-Core and a slower Mac 8-Core, in most cases the Quad will be the winner.
For example:
Mac Quad-Core 2.93 with 8GB 3.259,- US$
Mac 8-Core 2.93 with 8 GB 5.999,- US$
In most cases, if not more than 4 REAL cores are used, they both will be alike. In some cases, if really all cores are used, the 8-Core machine will be ~56% faster.
Is it worth the difference? Yes, but only if you really need it.
The memory limitations might hurt some of you but if you are doing normal stuff, 8GB should at least do it for the next 2 years. Also I guess 4Gb modules will work and should also become much cheaper in the next years.
This should be proofed, though.
But over all, I agree, all those Mac Pro machines are too expensive, also thinking about the current world economic situation.
Sorry for my bad English - I am Austrian
Anybody know when 12GB DIMMs will be available? 12x8!
Never. 8GB DIMMs have surfaced and 16GB will come sometime this year. It still hasn't been confirmed if Registered DIMMs will work in the Mac Pro (they should) and unbuffered DIMMs are likely to only come in sizes up to 4GB.
4GB modules are likely to be cheaper this round and more in line with the $ per GB price of 2GB modules.
Please read this doctoral dissertation from Cambridge University on the challenges of SMT scheduling, and tell me again that very little needs to be done.
Cool link. Thank you for sharing that. Note I'm not trying to argue with you, but i want to share some thoughts:
At first blush, I don't really see anything that refutes gnasher729's assertion that the change to the scheduler is small. The thesis of the paper is merely that SMT processors "can be used more effectively with an operating system that is aware of their characteristics." It doesn't seem to be claiming that the problem is necessarily difficult.
One interesting section is 4.7, where the author admits:
One might imagine that further refinements in Nehalem's implementation of SMT might make this caveat yet more applicable.
One other thought: back when OSX was leading a "secret double life" in the labs they were developing on Pentium 4 hardware. Indeed, that first round of developer hardware (the ones that had to be returned after the real Mac Pros were released) were just PowerMac aluminum cases with Pentium 4 hardware on the inside.
While this doesn't mean that they made OSX's schedule aware of SMT, it does push back the date at which they were hacking on SMT hardware to around the same timeframe when Bulpin submitted his dissertation. Moreover, Intel's roadmap has promised the reappearance of SMT in Nehalem for quite a while, so it's not like Apple hasn't seen the train coming down the tracks.
Nevertheless, my bet is that Leopard's scheduler was not SMT aware and that Snow Leopard will probably be the first shipping version to be so.
Thanks again for the link. That's my kind of geek pr0n.
That's possible - especially since Nehalem has more functional units to exploit, and the whole point of SMT is to avoid having idle functional units.
The real hard issues, however, remain. The scheduling is dynamic, and changes millisecond by millisecond. One millisecond you may have 8 threads running on 8 cores, next millisecond 16 threads running on 8 cores, and the next millisecond 8 threads running on 4 cores. Since moving a thread can be expensive, trying to keep it all "optimal" millisecond by millisecond by moving threads around can make it slower.
gnasher729's scenario, while reasonable, is a static analysis that misses the issues that dynamic thread scheduling exposes.
For more Geek Porn, look at Chip Multi Processing aware Linux Kernel Scheduler (pdf)
! Searching for "linux smt scheduler" will show that this area is an active one.Indeed, especially in the case we've been discussing here, which seems especially perverse to me: a host operating system apportioning a couple of virtual cores to another operating system in a VM. There are two different schedulers involved here, and who knows how they'll interact. Or maybe not. Hmm. Anyway, my head explodes thinking about it.
Some thoughts
There is a huge discussion about those new mac pro, so I will reassume the whole thread for the new readers:
Some people says the new MAC PROs are too expensive.
Some others says it's a huge step foward.
Some that it's a huge step backward.
Some says that the single chip edition sucks because of ram limit (8gb).
Some says that you won't need more that 8gb of ram in the next 2 years.
Some that the single 2.93 chip is the only one that gives you "bang for your bucks".
Some that only the most expensive octo 2.93 is the one that is really enough powerful to be taken in consideration.
Some says to wait another year for the 6 cores edition because this is a useless update and the next one will be much better.
Some that we will see the real advantage only when snow leopard comes.
Some that we can already see huge advantage on the programs that uses multi-core processing.
But some says that there are no program that really uses multi-core advantage. Some others don't.
Some says that hyperthreading and ecc-ram is super good.
Some that hyperthreading can be an obstacle that lower the performances and ecc-ram is useless because you don't really need it.
Some say dell are much better than apple.
Some says that apple is loosing touch.
Some other give kudos to apple for having released one of the world's most powerful computer.
Many different thoughts for a computer that is only 10 days old and that only few people actually used (or even saw)... Not to say that apple is the only one who released those xeon nehalem chips by now...
So, right now, after I red the whole thread, i'm pretty sure of this:
If you have the need and the money, grab what you can.
And if you don't have the need and the money, don't grab what you can't.
Do it.
But don't do it.
Think about it.
No don't.
Yes.
No.
?
.
There is a huge discussion about those new mac pro, so I will reassume the whole thread for the new readers:
Some people says the new MAC PROs are too expensive.
Some others says it's a huge step foward.
Some that it's a huge step backward.
Some says that the single chip edition sucks because of ram limit (8gb).
Some says that you won't need more that 8gb of ram in the next 2 years.
Some that the single 2.93 chip is the only one that gives you "bang for your bucks".
Some that only the most expensive octo 2.93 is the one that is really enough powerful to be taken in consideration.
Some says to wait another year for the 6 cores edition because this is a useless update and the next one will be much better.
Some that we will see the real advantage only when snow leopard comes.
Some that we can already see huge advantage on the programs that uses multi-core processing.
But some says that there are no program that really uses multi-core advantage. Some others don't.
Some says that hyperthreading and ecc-ram is super good.
Some that hyperthreading can be an obstacle that lower the performances and ecc-ram is useless because you don't really need it.
Some say dell are much better than apple.
Some says that apple is loosing touch.
Some other give kudos to apple for having released one of the world's most powerful computer.
Many different thoughts for a computer that is only 10 days old and that only few people actually used (or even saw)... Not to say that apple is the only one who released those xeon nehalem chips by now...
So, right now, after I red the whole thread, i'm pretty sure of this:
If you have the need and the money, grab what you can.
And if you don't have the need and the money, don't grab what you can't.
Do it.
But don't do it.
Think about it.
No don't.
Yes.
No.
?
.
ECC modules are almost always also Registered. *ESPECIALLY* for high-capacity. If you go to crucial.com and ask for memory for the new Mac Pro, though, it does only show Unbuffered.
I know...
I guess the opinions will get less and less clear... As for me, I had to buy a mac pro to get in the buisness, so the new octo 2.66/16gb ram/radeon 4870HD is a real deal, considering I come from a Macbook pro core duo 2ghz e 2gb of ram
DP octo 2.26 or the SP quad 2.66??
So...(i) I'm ready to buy, and (ii) have read through this fascinating thread. My basic question is this (and some posts have touched on this): am I better off with the DP octo 2.26, or the SP quad 2.66 (or SP quad 2.93 for that matter)?
(One thing I DO know is that whichever one I get will blow my current PowerMac G5 DP 2.5GHz right out of the water - but that's besides the point.)
A little background: I do audio work on my current G5 (Logic Studio, samples, plug-ins, etc.), and to a lesser extent, some video work using FCE. I don't know whether Logic is already or will be optimized for multi-cores and/or multi-processors. I just want something that will last me a good 6-7 years.
So boys and girls: more cores or higher clock speed?? Any suggestions or comments are greatly appreciated!
So...(i) I'm ready to buy, and (ii) have read through this fascinating thread. My basic question is this (and some posts have touched on this): am I better off with the DP octo 2.26, or the SP quad 2.66 (or SP quad 2.93 for that matter)?
(One thing I DO know is that whichever one I get will blow my current PowerMac G5 DP 2.5GHz right out of the water - but that's besides the point.)
A little background: I do audio work on my current G5 (Logic Studio, samples, plug-ins, etc.), and to a lesser extent, some video work using FCE. I don't know whether Logic is already or will be optimized for multi-cores and/or multi-processors. I just want something that will last me a good 6-7 years.
So boys and girls: more cores or higher clock speed?? Any suggestions or comments are greatly appreciated!
Well, from my point of view, considering the future development of the os, the longevity of the investment you want to do, I would go for a Xeon Nehalem Octo 2.26.
Why?
- It's upgradable (as a dual socket motherboard. As far as I know, the single processor units don't have a dual socket motherboard)
- No memory caps (RAM should be useful to you, considering you might work with huge audio libraries).
- The more we'll go through software development, the more the multi-core power will be used and win against single raw power of one proc.
Please read this doctoral dissertation from Cambridge University on the challenges of SMT scheduling, and tell me again that very little needs to be done.
I skimmed through it, and I tell you again, very little needs to be done.
A processor using hyperthreading can be described very easily: It pretends to have twice the number of cores; these virtual cores come in pairs so that using both virtual cores of a pair reduces the performance to some value like 60%. So all the OS needs to know is that both virtual cores from a pair should only be used if there is no pair of cores that is unused. This is a rather trivial change in the scheduler.
Apple has more complicated scheduling algorithms already implemented in Leopard.
I skimmed through it, and I tell you again, very little needs to be done.
A processor using hyperthreading can be described very easily: It pretends to have twice the number of cores; these virtual cores come in pairs so that using both virtual cores of a pair reduces the performance to some value like 60%. So all the OS needs to know is that both virtual cores from a pair should only be used if there is no pair of cores that is unused. This is a rather trivial change in the scheduler.
Apple has more complicated scheduling algorithms already implemented in Leopard.
I can say that Vista is quite stupid in its scheduling. If you have one thread that wants 100% of one CPU core, Vista will bounce it continuously between all available cores in turn. So much for making any decent use out of your CPU cache. And if you have two threads, it will bounce both of them. Which means that often on an SMT CPU, you will have both threads on the same physical core, nearly halving each thread's speed.
I tried to run a distributed computing app on my Core i7, and told my app to use only four threads to match the four physical cores. If I use Vista's task manager to set CPU affinity, I can get full performance. If I let Vista bounce the threads around, I only get about 60% of max performance. (This app handles SMT fairly well; if I turn SMT off, I get 'x' performance. If I turn SMT on, I get 'x*0.95' performance. But for some of the active tasks I do, like video encoding, I get 'x*1.6' performance, so I want SMT on for the things I really care about.)
OSX already supports pthreads. I'm not sure what GC is supposed to do that's better, unless OSX's pthreads implementation is a poor one, in which case why not improve the pthreads implementation rather than bother with GC?
FunkyChicken said:
Both are "upgradeable" if you want to void the warranty.
The single cpu model can be "upgraded" to 3.20GHz (soon to 3.33GHz according to Intel's roadmap). And the dual-cpu model can also be upgraded to dual 3.20GHz.
"Upgrading" to a single 3.20/3.33GHz cpu will cost around $1,000.
"Upgrading" to dual 3.20GHz cpus will cost around $3,200 (2x$1,600).
FWIW, a single quad 2.93GHz with 8GB RAM costs $3249 from Apple. The dual-quad 2.26GHz with 8GB costs $3399 from Apple. Those are probably the two configurations that you should be looking at.
The main disadvantage, today, of a single cpu model is the RAM limitation, but this could be Apple saying that they don't support (have not tested) 4GB RAM sticks on the single cpu models. The memory controller on the cpus can handle more than 8GB of RAM, especially the Xeon ones used on the Mac Pro. The desktop ones (Core i7) can handle 24GB (6x4GB) of RAM.
But in the audio business, just like many more other businesses, what you want is a very stable computer, playing with things like cpus, is not the way to go. I would wait for real world benchmarks to see how both models compare in terms of capabilities with the apps I use the most (in this case Logic Pro and FCP).
If I was a "company" I would go for the more powerful one whatever the price. If I was a freelancer, I would choose the "best bang for the buck" (it could be the dual cpu model, or the faster single cpu model, we will see). Because, next year, in a couple of years and certainly in 6-7 years, there will be more powerful models at equivalent prices.
And before then Logic Pro and FCP will also have evolved so much that they may require (or take advantage of) some hardware that isn't currently offered on the new Mac Pro (special gpu or proprietary co-processor, vector unit, larrabee, or whatever). 6-7 years is a long time.
6-7 years ago, digital audio was still 16/24bits @ 44.1/48KHz stereo 48/64 tracks, today we have up to 32bits @ 192K, 7.1+ surround, almost infinite number (256/512) of tracks... What will we have in 6-7 years?
Even if you plan on keeping your next purchase for 6-7 years, you WILL want the new Mac xxx turbo, that takes advantage of Logic Pro v.X or FCP v.Y that will be available in the spring of 2012 (or so).
I agree with you, but when i was talking about upgradability, I was pointing the fact that the future is multicore, so i think it's pretty useless to upgrade a single core machine instead of a multi core.
The new 2009 Mac Pro is rubbish in my opinion... Its Price/Performance does not match..
Snow Leopard should fix a lot of the issues.
It won't fix the serious drop in single-threaded performance, which is what many posters are complaining about.
Single threated performance is going to be less and less important than multi core performance.
Everything is going multicore.