Mixing 8 and 16 GB RAM modules - dual vs triple channel

[Spacedust]

I just got 6x 16 GB Samsung PC3-14900R 1866 MHz RAM 2Rx4 for my Mac Pro 2012 (in order to move 4 of them to a new Mac Pro in future).

The funny thing is when I put my old 2x 8 GB Kingston PC3-10600R 1333 MHz 2Rx4 in slot 4 and 8 it also works together giving a total of 112 GB RAM. OWC says it's not possible, but actually it is.

I don't see any lower results when using Dual vs Triple Channel. Is there any benchmark clearly showing this ?

 

[flowrider]

OWC is sometimes out to lunch! I have learned over the years to:

1. Believe little of what they say!

2. Not to buy the stuff they sell!

Each X56xx CPU is capable of running 56GB of RAM on a Mac Pro. There are threads on this. Using RAM in a triple channel configuration is ideal, but there are been other posters here, who report no discernible difference in real life applications.

Lou

 

[IowaLynn]

I just got 6x 16 GB Samsung PC3-14900R 1866 MHz RAM 2Rx4 for my Mac Pro 2012 (in order to move 4 of them to a new Mac Pro in future).

The funny thing is when I put my old 2x 8 GB Kingston PC3-10600R 1333 MHz 2Rx4 in slot 4 and 8 it also works together giving a total of 112 GB RAM. OWC says it's not possible, but actually it is.

I don't see any lower results when using Dual vs Triple Channel. Is there any benchmark clearly showing this ?

2Rx4

Because both set are Registered, they work. Most of the time folks don't buy Registered 8GB DIMMs. 16GB DIMMs have to be Registered.

 

[Spacedust]

Each X56xx CPU is capable of running 56GB of RAM on a Mac Pro. There are threads on this. Using RAM in a triple channel configuration is ideal, but there are been other posters here, who report no discernible difference in real life applications.
Lou

That was related to a single CPU Mac Pro, dual is able to run up to 64 GB for each CPU giving a 128 GB total and who knows maybe even more (288 GB according to Intel specs).

OS X is limited to 128 GB by now.

 

[ZombiePhysicist]

I just got 6x 16 GB Samsung PC3-14900R 1866 MHz RAM 2Rx4 for my Mac Pro 2012 (in order to move 4 of them to a new Mac Pro in future).

The funny thing is when I put my old 2x 8 GB Kingston PC3-10600R 1333 MHz 2Rx4 in slot 4 and 8 it also works together giving a total of 112 GB RAM. OWC says it's not possible, but actually it is.

I don't see any lower results when using Dual vs Triple Channel. Is there any benchmark clearly showing this ?

I just went through something similar in this thread:

https://forums.macrumors.com/threads/1868343/

I ran geek bench with just 96GB of memory using 3 sticks per CPU, and then added a 4th. The geek bench scores were basically identical for single processor use, and showed about a 6.5% decrease in speed when using 4 sticks per CPU in multicore tests. The tests used Geekbench 3 in 64bit mode.

Give it a try and see.

Perhaps there is less of a slow down when using the smaller memory size in slot 4 because it's some how less likely to access that memory? Not sure, this memory setup is weird.

 

[Spacedust]

I'm staying with 96 GB RAM ! Why see memory results:

112 GB RAM on the left, 96 GB RAM on the right

Full results:

96 GB RAM: http://browser.primatelabs.com/geekbench3/2513819

112 GB RAM: http://browser.primatelabs.com/geekbench3/2513879

AES score has doubled with Triple Channel !!!

Also very intersting thing is that Geekbench shows my Mac Pro 2010 as 2012 model because it was actually built in March 2012 (only serial is from 2010).

I wonder if someone rich has gotten 32 GB modules to get 6x32 GB = 192 GB - example for Windows 7 usage The CPU's itself are able to run up to 288 GB.

http://www.bhphotovideo.com/c/produ...ing_owc1333d3z3m128_128gb_32x3_kit_sdram.html

 

[benjaprud]

I have read an article which compared the performance between several unbalanced memory configurations on Nehalem/Westmere Xeons. The impact on performance is different if you use two or four identical modules or modules with different capacities. I'll give you the link if I find it.

EDIT : I finally found the document from IBM here, it's a great summary of memory performance and best practices on different setups. Here is another article (the most complete in the lot), it explains a lot about different types of memories and the best practices on Nehalem/Westmere Xeons to achieve best bandwidth, latency or capacity. If you haven't had enough you can read more here from Fujitsu (interesting reading too), here from Kingston and there.

After reading all these papers I have drawn the following conclusions about the best possible memory configuration for Nehalem/Westmere MP and the traps to avoid :

• Populating slots : Putting 3 identical DIMMs on the three slots the farther from the CPU is the best possible configuration for memory bandwidth, putting 4 forces 2-way interleave and has a small impact (24% on bandwidth, 0 to 15% on application performance, 1 to 5% on a mixed usage) but some applications benefit more from increased memory capacity, do it only if you need the extra capacity. Putting less than 3 is bad (gets you 1/3 or 2/3 of achievable bandwidth and is what we call single or dual-channel, single-channel has a 20% average performance penalty).
  
  
• Mixing capacities : A configuration with all channels populated and same capacity across channels is considered "near-balanced" (for example in a MP 2GB - 4GB - 4GB - 2GB is near-balanced) with close to optimal performance. Mixing capacities in an unbalanced configuration forces 2-way interleave and has potentially the same impact on performance than putting 4 DIMMs (13 to 24% less bandwidth depending on your setup). The second link has guidelines to build near-balanced configurations as well as benchmarks comparison between balanced, near-balanced and unbalanced.
  
  
• Dual-CPU : On dual-CPU it is recommended to use the same memory configuration on both CPU for lower latency. The effect on performance is about 2-3% on a moderately asymmetrical setup and up to 20% if you populate only one CPU.
  
  
• Ranks : Dual-rank modules offer the best interleaving performance, however the difference with single or quad is small. Quad-rank will always run at 1066MHz maximum and 800MHz on 4 DIMMs setup. Quad-rank DIMMs are always registered, they also consume a bit less power. Higher ranks allows to put more memory chips on a DIMM allowing for higher capacities and/or cheaper/older tech DIMMs. On DDR3 ECC sticks, single-rank should have 9 memory chips, dual 18 chips and quad 36 chips. Single and dual-rank can be mixed together although it may hit performance a bit (resulting in near-balanced configuration). If quad-rank is mixed in the lot, other modules have to be registered.
  
  
• Voltage : 1.35V modules offer the same performance than 1.5V modules while consuming 20% less power. Quad-rank 1,35V memory will always run at 800MHz. On 4 DIMMs setup with 1333MHz capable memory and CPU it's better to use 1.5V powered modules as the MP's EFI will force the memory at 1066MHz if there is more than one DIMM per channel running at 1.35V (other UEFIs/BIOSs from other manufacturers can choose to force the memory at 1333MHz/1.5V). That was the answer to the original question that led me to all this research.
  
  
• Registered memory : Registered memory works but has a slightly higher latency (one cycle) and power consumption (about 0.75W/DIMM). It is meant for servers with lots of DIMMs (like 12 or 16) to lower the electrical load on the memory bus. It is also used on quad-rank high capacity modules (16 and 32 GB). It's usually a bit more expensive too. It cannot be mixed with unbuffered memory.
  
  
• Mixing different types of DIMMs has a sometimes unpredictable (given the number of possible combinations) and potentially important impact on performance/compatibility and is to be avoided when possible. If you're concerned about performance it's up to you to determine the impact on your workflow. Some applications will be more sensitive to bandwidth or latency while others will be more sensitive to capacity. The more powerful the processor, the greater the influence of the memory parameters.

This post has been repeatedly and heavily edited since its first writing and might have contained inaccuracies and misleading information. It should be accurate by now (as far as the linked documentations are).

 

[flowrider]

Also very intersting thing is that Geekbench shows my Mac Pro 2010 as 2012 model because it was actually built in March 2012 (only serial is from 2010).

I don't think that's the reason. What CPU is in your machine? Geekbench keys off the CPU for identification of a 5,1 MacPro. I'm running X5677s in my DP 5,1 and Geekbench only id's my machine as a MacPro5,1 with no id year because the X5677 was never used by Apple. My machine was built in 2011.

Lou

 

[ActionableMango]

After reading all these papers I have drawn the following conclusions about the best possible memory configuration for Nehalem/Westmere MP and the traps to avoid :

Thanks for putting that all in one place.

 

[666sheep]

If quad-rank is mixed in the lot, it has to be put in the furthest slot from the CPU (and other modules have to be registered).

This one has made me curious and it turned out that it doesn't have.
I put one Quad ranked in a slot 3 in mix with 2 Dual ranked ones in slot 1 & 2, and in other way I put one Dual ranked in slot 2 with 2 Quad ranked in slots 1 & 3. Both combination worked OK. That was in single CPU W3670 one, so only 1066 MHz ofc. Dual ranked sticks were 12800s, quad ranked 8500s, all 16GB ones so registered. Just for the record.

 

[benjaprud]

I'll put that one off then. These info are extracted from intel's specs and were supposed to be accurate in mid-2009, DDR3 technology may have evolved since. There can also be some differences from one computer manufacturer to another. Thanks for trying, one would need loads of RAM sticks to try all the configurations described. I haven't verified any of this myself as I always had a balanced setup, any feedback is welcome.

 

[AidenShaw]

I'm staying with 96 GB RAM ! Why see memory results:

112 GB RAM on the left, 96 GB RAM on the right

Good choice if your main application is Stream.

Not as important if you run useful applications.

Stream is a synthetic benchmark written to bypass all levels of cache and show raw memory system performance.

Useful applications are usually written to take the best advantage of cache, and often mask raw memory system speeds.