4GB + 4GB + 4GB + 16GB? Mix 'n' match UDIMMs and RDIMMs?

Torley

macrumors member
Dec 3, 2008
37
0
0
So, I know this Apple article: http://support.apple.com/kb/HT6064 states not to mix UDIMMs and RDIMMs. Looks like all the 16GB modules are RDIMMs. Apple's own recommendations tend to be more stringent than what's tested out in the wild.

Has anyone tried adding a 16GB to a 12GB (4GB + 4GB + 4GB) base config for a total of 28GB, and if so, what are your findings? Or is it preferable to play it safe and add two 8GB UDIMM modules (replacing one 4GB one) for a total of 24GB?
 

VirtualRain

macrumors 603
Aug 1, 2008
6,304
114
0
Vancouver, BC
So, I know this Apple article: http://support.apple.com/kb/HT6064 states not to mix UDIMMs and RDIMMs. Looks like all the 16GB modules are RDIMMs. Apple's own recommendations tend to be more stringent than what's tested out in the wild.

Has anyone tried adding a 16GB to a 12GB (4GB + 4GB + 4GB) base config for a total of 28GB, and if so, what are your findings? Or is it preferable to play it safe and add two 8GB UDIMM modules (replacing one 4GB one) for a total of 24GB?
I did a quick search and while most people heed the warning not to mix UDIMMs and RDIMMs (which all server vendors caution against) there's at least one person who tried it without success. So I'd probably give up on that. I'm not sure what kind of performance you'll get with mixed DIMM capacities either... It could default to single channel operation which would suck. I'd really go for matched DIMMs (16, 32, or 64). They're not that pricey... http://www.superbiiz.com/query.php?dp=1&dt=2&categry=727&name=DDR3-SDRAM-Server-240-Pins&brand=&pa0=&pa1=DDR3-1866&pa2=&pa3=&pa4=&pa5=&pa6=&pa7=&stock=No&nl=30&searchStr=Search+from+current+results&ob=r&myanchor=#displaytop
 

Torley

macrumors member
Dec 3, 2008
37
0
0
Thanks — do you have a pointer to the performance drop % in single channel mode? If it's small or negligible in real-world performance I can live with that. Otherwise, right now, I'm eyeing OWC and and that page you linked to.
 

longname

macrumors newbie
Feb 1, 2014
15
1
0
I don't think you can mix between UDIMM and RDIMM. According to my knowledge, RDIMM RAMs run one clock cycle behind he equivalent unregistered DRAM. RDIMMs will have lower latency and better bandwidth than UDIMMs.
This two types of RAM run differently, so my suggestion is not to run them together.
I already ordered a 16GB registered RAM. It will arrive on Tue. My system has the same 3x4GB with one non used slot! So I will post the result for you!
 

longname

macrumors newbie
Feb 1, 2014
15
1
0
Luckily, I found a registered RAM in my case, so I could test it out!
My result is the system didn't boot when I mixed registered and unregistered RAMs together. The system booted normally when I pulled either UDIMM or RDIMM RAM out.


 

Torley

macrumors member
Dec 3, 2008
37
0
0
Very grateful for your photographic example and research, longname!

Hope this helps out future people looking for the same thing. Thanks!
 

AidenShaw

macrumors P6
Feb 8, 2003
18,047
4,127
0
The Peninsula
carpet?

Luckily, I found a registered RAM in my case, so I could test it out!
My result is the system didn't boot when I mixed registered and unregistered RAMs together. The system booted normally when I pulled either UDIMM or RDIMM RAM out.

Image
Image
Just a note that it's very risky to put electronic components on carpet - static electricity can easily destroy the circuits.
 

GoogleFish

macrumors newbie
Aug 12, 2010
18
0
0
I did a quick search and while most people heed the warning not to mix UDIMMs and RDIMMs (which all server vendors caution against) there's at least one person who tried it without success. So I'd probably give up on that. I'm not sure what kind of performance you'll get with mixed DIMM capacities either... It could default to single channel operation which would suck. I'd really go for matched DIMMs (16, 32, or 64). They're not that pricey... http://www.superbiiz.com/query.php?dp=1&dt=2&categry=727&name=DDR3-SDRAM-Server-240-Pins&brand=&pa0=&pa1=DDR3-1866&pa2=&pa3=&pa4=&pa5=&pa6=&pa7=&stock=No&nl=30&searchStr=Search+from+current+results&ob=r&myanchor=#displaytop
So if you can't add a 16gb stick to a stock 12gb config, based on the price of a 16gb stick as linked above, seems to make sense ($ wise) to buy a single 16gb stick and pull the 3x4gb oem sticks to get a 16gb config - this based on the cost from apple to add a single 4gb stick (about $100). Gives you an opportunity to upgrade as needed/prices of 16gb sticks comes down?
 

AidenShaw

macrumors P6
Feb 8, 2003
18,047
4,127
0
The Peninsula
So if you can't add a 16gb stick to a stock 12gb config, based on the price of a 16gb stick as linked above, seems to make sense ($ wise) to buy a single 16gb stick and pull the 3x4gb oem sticks to get a 16gb config - this based on the cost from apple to add a single 4gb stick (about $100). Gives you an opportunity to upgrade as needed/prices of 16gb sticks comes down?
Yes, that sounds reasonable. Especially since in a few months 32 GiB DIMMs will probably be priced around where 16 GiB DIMMs are now.

Many tests have shown that most workloads run nearly as fast with fewer channels of memory. That 10 MiB L3 cache helps hide the small loss of main memory speed from most apps.
 

peabo

macrumors regular
Feb 14, 2008
199
0
0
I know that with 2x16GB DIMMs, my geekbench score was 31000 but with 4x16 it was over 33000, so it does make some kind of difference. Having 4x4 would likely be faster than having 1x16.
 

AidenShaw

macrumors P6
Feb 8, 2003
18,047
4,127
0
The Peninsula
I know that with 2x16GB DIMMs, my geekbench score was 31000 but with 4x16 it was over 33000, so it does make some kind of difference. Having 4x4 would likely be faster than having 1x16.
The attached chart from Fujitsu (http://globalsp.ts.fujitsu.com/dmsp/Publications/public/wp-sandy-bridge-ep-memory-performance-ww-en.pdf) describing Sandy Bridge performance with one, two and four channels shows just a few percent loss with one channel on SPECint.

Your 31000/33000 for two to four is about 6% - but it's a twelve core, right?

A quad core is much less thirsty for bandwidth - they're "basic" processors.

Even so, 6% means that a two hour render would take two hours and seven minutes. Hardly a "make or break" difference.

Edit: I mis-read the paper, the table is for number of ranks, number of channels.
 

Attachments

Last edited:

AidenShaw

macrumors P6
Feb 8, 2003
18,047
4,127
0
The Peninsula
Is this true? :confused:

I know Apple prices for ram is normally high, but $100 for 4gb seems excessive.
It's true, but you'd probably never notice the difference in speed on your workflow.

A few percent on memory-intensive tasks. 0% on cache-friendly tasks.

I think that your idea of getting a 16 GiB DIMM now, and a second when the price stabilizes is good.

It makes more sense than wasting $100 on the 4 GiB when you upgrade later. (And maybe you'd be able to get nearly $300 selling the three 4 GiB DIMMs to other people who bought the base....)
 
Last edited:

Torley

macrumors member
Dec 3, 2008
37
0
0
Great conversation here, again, I really hope this will help future nMP RAM upgraders. I like the notion of replacing my 3x4GB with 1x16GB for now and expanding accordingly in the future, seems better for long-term.
 

AidenShaw

macrumors P6
Feb 8, 2003
18,047
4,127
0
The Peninsula
Great conversation here, again, I really hope this will help future nMP RAM upgraders. I like the notion of replacing my 3x4GB with 1x16GB for now and expanding accordingly in the future, seems better for long-term.
Thanks.

And I like that HP lets me order ProLiants with 0 GiB of RAM. I don't have to overpay for DIMMs that I'll put in the eWaste bin.

You shouldn't be forced into a small set of memory configs when you order the system.
 

JQuick

macrumors member
Feb 2, 2014
77
0
0
The attached chart from Fujitsu (http://globalsp.ts.fujitsu.com/dmsp/Publications/public/wp-sandy-bridge-ep-memory-performance-ww-en.pdf) describing Sandy Bridge performance with one, two and four channels shows just a few percent loss with one channel on SPECint.

Your 31000/33000 for two to four is about 6% - but it's a twelve core, right?

A quad core is much less thirsty for bandwidth - they're "basic" processors.

Even so, 6% means that a two hour render would take two hours and seven minutes. Hardly a "make or break" difference.
I suggest that you reread that white paper. You have misinterpreted it.

The table you cited from page 16 of that report does not show "just a few percent loss on one channel". That data is comparing interleaving among ranks within a channel. Interleaving among channels is much different thing, which is discussed on page 14 of that paper. I will quote from it here:

There may be good reasons for 2-way and 3-way interleaving with a moderate loss in performance: a low memory capacity that is needed or minimization in the number of DIMMs in order to save energy. We advise against 1-way interleaving, which is not strictly speaking interleaving and is only referred to as such for the sake of the systematics involved. In this case, the performance potential of processors and memory system are not in a well-balanced relationship to each other.
Data on that page show Specint performance dropping by between 11%-26% by going from 4 to 1 channel operation. The more bandwidth intensive STREAM scores reflect the worst case 67%-71% drop in memory throughput.

I was wondering about this topic a week or so ago and compared some random geek bench memory numbers for 4-core 12 and 16GB models to see the difference between 3 and 4-way interleaving. I saw memory performance numbers about 10% higher for single core operations and 25% higher for multi-core when 4 channels were present.

Swapping 3x4GB for a single 16GB RDIMM seems like a very poor trade.
 

longname

macrumors newbie
Feb 1, 2014
15
1
0
Great conversation here, again, I really hope this will help future nMP RAM upgraders. I like the notion of replacing my 3x4GB with 1x16GB for now and expanding accordingly in the future, seems better for long-term.
Yeb!! I'm doing the same thing! I will resell all of my 4GB RAMs, and get another 16Gb
 

JQuick

macrumors member
Feb 2, 2014
77
0
0
Excellent. Thanks for posting this.

The relevant data for comparison is the multicore memory performance which most cleanly isolates the 2x vs. 4x memory interleaving. I consider the single core figure irrelevant since memory performance is critical to so many other aspects of the system. Filesystems, network buffers and data flowing to/from GPUs, etc. all rely heavily on the memory subsystem. Even if the current primary workflow is CPU bound on a single thread other parts of the system may access memory at the same time.

Calculating the difference in throughput (5181-3734)/3734 shows the 4x16 configuration is 39% faster in this test.

It is clear that the system will boot and operate with anywhere from 1-4 channels populated. It is also clear that the impact on memory subsystem performance is significant when operating with fewer than 4 DIMMs. As a temporary or cost conscious measure I think 3-way interleaving may still provide adequate performance in less bandwidth intensive workflows (about 25% boost to memory bandwidth by adding the fourth DIMM). Obviously Apple does as well. But the 2-way hit is so significant that I can see why Apple discourages it (by not listing it as a supported configuration at all).

This saved me a lot of legwork to determine my eventual memory upgrades after my own system arrives. Thanks again.

n.b.
I initially implied there was a 25% drop in performance by going from 4 to 3 DIMMS, when it is a 25% boost going from 3 to 4 DIMMS. Oops, mathing. It is clearly not linear but still highly significant.
 
Last edited:

spaz8

macrumors 6502
Mar 3, 2007
407
8
0
Maybe I'm missing something, but looking at the final geekbench numbers in single and multi-core looks like there is only a 3-5% difference in 2 vs. 4 channels used.

Clearly in memory performance there is a difference.. but it looks like its about a 4% difference conceivable by the user.
 

VirtualRain

macrumors 603
Aug 1, 2008
6,304
114
0
Vancouver, BC
Maybe I'm missing something, but looking at the final geekbench numbers in single and multi-core looks like there is only a 3-5% difference in 2 vs. 4 channels used.

Clearly in memory performance there is a difference.. but it looks like its about a 4% difference conceivable by the user.
Yeah, It really depends on how much of the dataset can reside in cache... and these days the huge cache attached to most CPU's will mask any issues in the underlying memory architecture.

Having said this, I believe most people in this forum would be lying if they said they didn't care about 5% performance. That's the difference a CPU generation can make.

I really think people are crazy if they're going to buy a nMP and outfit it with a single 16GB DIMM. Man up and populate all those DIMM slots! :rolleyes: