Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

oban14

macrumors 6502a
Original poster
Jan 4, 2008
554
1
I read in a thread here that for the fastest performance, you should have 8 identical ram sticks, and the second choice is to have 4.

My Mac Pro is coming with two sticks (1GB each) and I ordered another four (2 gb each) for a total of 10 gigs of ram.

Am I really going to take a performance hit here? Would I be better off pulling the two 1 GB sticks and just having 8 GB of ram instead of 10 GB? :confused:
 
You will get a performance hit of around 10% or so. The reason is, by having 4 and 8 matching pairs it allows the Mac PRo to run at quad/8 channel mode which gives a performance boost in the writes to and from memory. By having a non-matching amount, that prevents that and forces it back into the standard dual channel mode.
 
I think this config is ok for the most part if you follow the instructions posted on Bare Feats.

http://www.barefeats.com/harper3.html

Thanks, bookmarked it. Puzzling that they needed to put the two 1 GB sticks on the top riser to work... would have probably kept me stuck for hours.

Thanks!

You will get a performance hit of around 10% or so. The reason is, by having 4 and 8 matching pairs it allows the Mac PRo to run at quad/8 channel mode which gives a performance boost in the writes to and from memory. By having a non-matching amount, that prevents that and forces it back into the standard dual channel mode.

It sounds like I might be better off pulling the two 1 GB Dimms then. I guess I could always sell them on fleabay.
 
Also note, that's a theoretical performance hit of that amount. In terms of real life performance hit, not so much. And, because of how OS X(and yes Vista too) handles memory, the impact is far less when you have a large amount of memory as the operating system will try and limit writes to and from memory.
 
Also note, that's a theoretical performance hit of that amount. In terms of real life performance hit, not so much. And, because of how OS X(and yes Vista too) handles memory, the impact is far less when you have a large amount of memory as the operating system will try and limit writes to and from memory.

Well, sounds like I might need to do some benchmarking when I get this beast. Should be fun. :D
 
yeah, i gotta say, this whole "all memory has to match" and everything seems really blown out of proportion.

i just find it hard to believe that adding more ram is ever going to actually make you go slower. (assuming it's installed correctly).
 
Have you considered adding 2 1GB chips to have 2 sets of matching 4s? I've ordered a set of 4 4GB chips, and then 2 1GB chips to match the Apple 2 1GB chips to have a second set of 4 chips matched.
 
OK, guys. Please define "Matched"

barefeats.com said:
Some of you just want to buy two 2GB FB-DIMMs to add to the two factory 1GB FB-DIMMs for a total of 6GB. That configuration will slow you down. Don't do it. Again, the BEST performance setup is eight matched FB-DIMMs. The second best is four matched FB-DIMMs.

OK, does "matched" mean that all the DIMMs are identical, or that just the pairs are identical? Can you have 1 matching pair from one manufacturer and 3 matching pairs from another manufacturer (giving you a total of 8 DIMMs)???

The reason why I ask is because I have the stock 2 x 1GB DIMMs coming in the machine from Apple. I want to upgrade the machine to a TOTAL of 8 1GB DIMMs. Is it OK to purchase 6 more 1 GB DIMMs and install them accordingly, or do all DIMMs need to "match" each other, same manufacturer?

Just seeing if I can save myself $128 bucks.
 
for DUAL channel to work (a requirement on a mac pro, incidentally), you need matched PAIRS.

for QUAD channel to work (not a requirement on a mac pro, but an option), you need matched SETS of either 4 or 8, which obviously are multiples of 4.

Now, back in the old days of single vs. dual channel DDR RAM, I did some tests and benchmarks of memory performance and system performance i general with my PC that had a dual-channel BIOS setting that could be easily adjusted.

I had a matched pair in the machine, and ran it on dual-channel mode, and compared to the same setup with single-channel mode forced, it did 5-10% better in memory-specific sub-tests in benchmarks. If you look at the entire system's performance, even in a memory-hungry and memory-intensive benchmarking app, you'll find that the actual system performance increase is less than 1%, all other things being equal.

Actual everyday system performance will be even less of an increase.

If you are going to need more than 14GB of RAM, then it makes sense to pull out the 1GB chips and replace them with all-matching RAM. Otherwise I don't think you'll ever see a real difference, unless the work that you do is writing/debugging memory benchmark software. That's probably not any of us. For most of us, having more RAM is more important than having marginally faster RAM. running in quad-channel doesn't double the speed or anything.

I generally trust BareFeats when they show me numbers.

I just went through their page on this (thanks for the link) and I must say I don't see any real numbers. Just anecdotes and 'someone I know told me it would be slower that way.'

This is of course directly contradictory to the earlier statement on the same page that you should use as much RAM as possible in your Mac Pro.

Sounds to me like you should buy 6 2GB chips (best value for the byte at the moment) and insert them in there with your 2 1GB chips that apple gives you. since more=better, then 14≥12, right?

sure, 16≥14, but you can save a few hundred bucks and not be stuck with two sticks of otherwise-useless ECC DDR2. I don't deal with ebay, so I would just have to let it sit around. No thanks!

And to answer the question about matched pairs, you don't have to match the manufacturer! Just make sure they are the same specs and size, and they're matched. Some people get weird about matching pairs, but honestly, the memory controller is handling everything anyway, and most RAM with the same specs and timings is equivalent these days...especially in ECC-land.
 
yeah, i gotta say, this whole "all memory has to match" and everything seems really blown out of proportion.

i just find it hard to believe that adding more ram is ever going to actually make you go slower. (assuming it's installed correctly).

Wrong! That's because you never use that much of memory you have.;)
 
How will 2 x 2gb dimms in addition to the 2 x 1gb dimms for a total of 6gb slow you down? I don't get it. I have a single quad core xeon Mac Pro and that is the setup I am running.
 
you never use that much of memory you have

Have you been using OS X for a while?

It will use up all of your RAM quite happily to avoid paging. I've got Photoshop, Illustrator, Acrobat, InDesign, Opera, Entourage, Word, iTunes, and Excel open right now, and my Mac Pro, with 6 GB of RAM, has less than 100mb of "Free" RAM. 4.2 GB of it is "inactive" according to iStatPro, and that's because I only have 3 windows open right now, and none of the big design apps has anything open on it. But when I click on inDesign, it will pop up in less than a second ready for my input, because it's not swapping anything out to the hard drive. Sometimes, I get the urge to get some more RAM, because I'm working on very heavy files and I can hear the hard drive grinding as I scroll around, which is all stuff that should be hanging out in the RAM, but can't because I don't have enough. If I had 12 or 16gb, my work flow would probably never page out, or at least very rarely. If I were in video production or animation or something, I can see how even 16gb wouldn't be enough to avoid paging.

on a scale of 1 to 1000, with quad-channel RAM at 1000 and paging from a RAID 0 array of 7200 RPM hard drives at 1, dual-channel RAM is a 990.

If that extra 2GB keeps you from paging, it's worth much more of your time than worrying about how many virtual memory channels there are currently active.
 
yeah, i gotta say, this whole "all memory has to match" and everything seems really blown out of proportion.

i just find it hard to believe that adding more ram is ever going to actually make you go slower. (assuming it's installed correctly).

I think that's a confusing way to look at it. Rather, look at it this way: having matched pairs could give you a performance boost. If you don't have matching pairs, you don't get that performance boost; but will still obviously get the benefit of of having the extra RAM.

(Of course though, extra RAM doesn't really make things faster - more accurately insufficient RAM makes things slower. Once you have enough, extra RAM makes no performance difference.)
 
Have you considered adding 2 1GB chips to have 2 sets of matching 4s? I've ordered a set of 4 4GB chips, and then 2 1GB chips to match the Apple 2 1GB chips to have a second set of 4 chips matched.

I think I'm gonna do that.

If there are tests on this RAM setup vs the 4x2 + 2x1 I'd really like to see them.
 
for DUAL channel to work (a requirement on a mac pro, incidentally), you need matched PAIRS.

for QUAD channel to work (not a requirement on a mac pro, but an option), you need matched SETS of either 4 or 8, which obviously are multiples of 4.

Now, back in the old days of single vs. dual channel DDR RAM, I did some tests and benchmarks of memory performance and system performance i general with my PC that had a dual-channel BIOS setting that could be easily adjusted.

I had a matched pair in the machine, and ran it on dual-channel mode, and compared to the same setup with single-channel mode forced, it did 5-10% better in memory-specific sub-tests in benchmarks. If you look at the entire system's performance, even in a memory-hungry and memory-intensive benchmarking app, you'll find that the actual system performance increase is less than 1%, all other things being equal.

Actual everyday system performance will be even less of an increase.

If you are going to need more than 14GB of RAM, then it makes sense to pull out the 1GB chips and replace them with all-matching RAM. Otherwise I don't think you'll ever see a real difference, unless the work that you do is writing/debugging memory benchmark software. That's probably not any of us. For most of us, having more RAM is more important than having marginally faster RAM. running in quad-channel doesn't double the speed or anything.

Cool. Thanks for the response. That helps clarify things. So, essentially there's no real-world significant speed increase from running things in QUAD channel mode. If there is, it's only marginal. Right?

Also, is there anything to the logic that you need "at least 1GB of RAM per Core?"
 
So will 4 x 1gb sticks be good?

I don't need more than 4 gb just yet. I am in a financial constraint and would rather have 4 gb now and add another 4 (2 x 2) later.

Will having 4 sticks of 1 gig be quick? Or is this much slower than 2 gb dimms?
 
I think that's a confusing way to look at it. Rather, look at it this way: having matched pairs could give you a performance boost. If you don't have matching pairs, you don't get that performance boost; but will still obviously get the benefit of of having the extra RAM.

(Of course though, extra RAM doesn't really make things faster - more accurately insufficient RAM makes things slower. Once you have enough, extra RAM makes no performance difference.)

Thanks for that post; I think that's a really good clarification that most people probably know, but that bears emphasizing.
 
The next question I have is, to run in QUAD channel, can you do this...

4 x 1GB
4 x 4GB

...and still benefit from the QUAD channel, or does having 4GB in one QUAD and 16GB in another create deficiencies that would drop the Pro back down to DUAL channel?
 
The next question I have is, to run in QUAD channel, can you do this...

4 x 1GB
4 x 4GB

...and still benefit from the QUAD channel, or does having 4GB in one QUAD and 16GB in another create deficiencies that would drop the Pro back down to DUAL channel?

I'm going to quote Barefeats:

"Some of you ordered your Mac Pro with the basic two factory 1GB FB-DIMMs. You are asking if you order four 2GB FB-DIMMs if you can install them along with the factory memory (for a total of 6 FB-DIMMs or 10GB of RAM).

I tested the 10GB scenario. I put four 2GB FB-DIMMs -- two on each riser in slots 1 and 2. Then I put the two 1GB FB-DIMMs in slots 3 and 4 on the top riser. With that config I get the same fast memory fill rates in Xbench as I did with just the four 2GB FB-DIMMs (two on each riser, slots 1+2). But I'm told by those who l know that the four matched FB-DIMMs will perform better when you get into a heavy duty multi-threaded app that uses a signficant amount of memory. Eight matched is even better.

The BEST performance setup is eight matched FB-DIMMs. The second best is four matched FB-DIMMs."
 
I'm going to quote Barefeats:

With that config I get the same fast memory fill rates in Xbench as I did with just the four 2GB FB-DIMMs (two on each riser, slots 1+2).

sounds like that's the money quote right there.....

But I'm told by those who l know that the four matched FB-DIMMs will perform better when you get into a heavy duty multi-threaded app that uses a signficant amount of memory. Eight matched is even better.

who are these people that know, and where are the numbers to back this assertion up?

i have 1 X 1 on my top riser and 1 X 2 in the bottom one, and have yet to see anything that's compelling me to change that.....
 
I'm going to quote Barefeats:
"(...)The BEST performance setup is eight matched FB-DIMMs. The second best is four matched FB-DIMMs. (...)"
Well... To say the truth, I don't buy that. Eight matched FB-DIMMs means each memory channel is linked to two FB-DIMMs and that actually leads to a small latency penalty when accessing modules seated in slots 3 and 4. Therefore, the best way to obtain a total of 4 Gb, 8 Gb or 16 Gb is to only use four FB-DIMMs of the same size instead of eight.

In addition, I believe the requirement to match all FB-DIMMs modules is an urban legend. While it is true that the chipset configures itself with some specific characteristics taken from all installed modules (i.e. the less stringent requirement of some module is made common for all, especially latencies), that doesn't mean all DIMMs shall be identical to get the best performances. Furthermore, from a technical standpoint and as far as quad channel is concerned, there's no difference between a 4x1 + 2x1 setup (6 Gb) and a 2x2 + 2x1 setup. In both case, the first part of the memory map can be accessed using all four channels simultaneously, while the last part cannot.

All documents I've found so far talk about matched pairs, because the two channels supported by each branch (one branch per riser) are interlocked. A matched pair is made of two FB-DIMM modules of the same size, same latencies, and same architecture. Although that's easier to achieve by using modules of the same model/brand, one could match modules of different model/brand.
Back to the quad-channel mode, nothing suggest nor imply they are interlocked. The whole thing appears to be a dual 128-bit memory path, and not a (synchronized) single 256-bit memory path. This leads to question the inner working of the benchmark tools because multiple threads are likely to take advantage of a dual memory path (a single thread would see no difference between dual-channel and quad-channel modes).

At the end of the day, I've yet to find any evidence to back Barefeat's claims...
 
Have you considered adding 2 1GB chips to have 2 sets of matching 4s? I've ordered a set of 4 4GB chips, and then 2 1GB chips to match the Apple 2 1GB chips to have a second set of 4 chips matched.

Did you get the 4x4GB from OWC or Transintl? I asked because OWC cannot quote a shipping time for the 4GB modules just yet.
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.