Nehalem memory -- 6 is better than 8 sticks

Discussion in 'Mac Pro' started by barefeats, Mar 15, 2009.

  1. barefeats macrumors 65816

    barefeats

    Joined:
    Jul 6, 2000
    #1
    Last night Lloyd Chambers of DigLloyd.com tweaked his DigLloydTools app (DLT) that we use to test memory throughput. It's now more accurate. The bad news is that it showed that when we put 8 sticks of memory in the 8-core 2.26GHz Nehalem, our throughput for memory read/write (memmove) dropped by 1/3. It turns a triple channel memory bus into a dual channel memory bus. Arggghh.

    Specifically, in our test, the combined read/write throughput dropped from 9261MB/s to 6195MB/s when we went from 6x2GB to 8x2GB configuration.

    Now, don't panic. That doesn't necessarily affect real world app performance unless the particular app you are running is saturating the memory bus. Which apps saturate? I don't know yet. I'm running our complete real world test suite including Pro Apps and 3D Games in both the 12G and 16G config. If I find anything that's significantly slowed by the 8x2G config, I'll post it here as well as on Bare Feats.
     
  2. rylin macrumors 6502

    Joined:
    Aug 18, 2006
    #2
    It's also important to note that memory throughput isn't everything.
    If you're actively using 14GB+ of memory, you're better served by 16GB than by 12GB for the simple reason that swapping is even slower ;)
     
  3. Nik macrumors 6502

    Joined:
    Jun 3, 2007
    Location:
    Germany
    #3
    Do you also have a QuadCore model at hand for testing?
    The question is if 4GB Sticks can be used in these models!

    Keep up the good work! :)
     
  4. barefeats thread starter macrumors 65816

    barefeats

    Joined:
    Jul 6, 2000
    #4
    Agree. In our After Effects CS4 tests with Total Benchmark, it gobbled up 13GB when I had it in 16GB config. When running in 12G config, it only had 10GB available because 2GB is reserved for "other apps." So as you can guess, it ran the benchmark 3 seconds faster with 16GB (117 vs 120).

    It's a balancing act. Do I need memory capacity or speed or both?
    If 4G modules were not so expensive, the best way to go would be 6x4G = 24G.

    As for 4-core Nehalem with only 4 memory slots, OWC will be trying the 4G modules out in the 4-core as soon as they get some in stock (probably this week). If it works, they will announce it on their site and their blog.
     
  5. canonballs macrumors member

    Joined:
    Mar 8, 2009
    #5
    Now the logical thing to ask would be something like this: With 8 sticks, is there an automatic way to pretend that slots 4 and 8 "aren't there" when memory throughput is lower than 12GB? Do we know for sure MP does not already do that?
     
  6. barefeats thread starter macrumors 65816

    barefeats

    Joined:
    Jul 6, 2000
    #6
    We tested that scenario. The 4th and 8th slot drag down the whole bus.
     
  7. Umbongo macrumors 601

    Umbongo

    Joined:
    Sep 14, 2006
    Location:
    England
    #7
    Apple sure did cause confusion with these new models. 2008 or 2009, quad or octo, 6 or 8 memory modules. They had avoided such things in the past with things like not using 8 core processors until the 3GHz were available and only offering the 2.8GHz in single processor configurations. I wonder if those who made such decisions moved on or actually how high such decisions go.
     
  8. Fomaphone macrumors regular

    Joined:
    Jan 11, 2009
    #8
    i'm sorry but i don't understand what core counts and clock speed have to do with memory bus speed. care to explain to a relative noob?
     
  9. cube macrumors G4

    Joined:
    May 10, 2004
    #9
    Duh. Why don't these machines have 6 and 9 (or more possibly, 12) slots respectively?
     
  10. Flash SWT macrumors 6502

    Flash SWT

    Joined:
    Mar 14, 2009
    Location:
    Houston, TX
    #10
    How about 4x2GB, are memory speeds similar to 8x2GB since it is in dual channel mode? I'd love to see this benchmark added to the DigLloydTools test.
     
  11. sejanus macrumors regular

    Joined:
    May 3, 2005
    #11
    I reckon for the octo models, get 12gb now (6x2) and later when the 4gb chips are better priced go to 24gb (6x4)

    Thats my plan anyway.
     
  12. Flash SWT macrumors 6502

    Flash SWT

    Joined:
    Mar 14, 2009
    Location:
    Houston, TX
    #12
    I'll be going with 4x2GB from Apple and then upgrading with an additional 2x2GB modules (when OWC gets around to offering that option) for a total of 6x2GB. I'm curious what kind of speeds I'll be seeing until then.
     
  13. -js- macrumors regular

    -js-

    Joined:
    Jul 30, 2008
    Location:
    Southern California
    #13
    Lovely! I can't wait to find out! And thanks for posting about the 1/3 drop in speed when using all 4 slots (or all 8 slots). Presumably if you used up 4 slots of one CPU, and then only 2 slots of the other CPU (in the octo, of course), then the speed would still drop, right? Just curious.
     
  14. VirtualRain macrumors 603

    VirtualRain

    Joined:
    Aug 1, 2008
    Location:
    Vancouver, BC
    #14
    Are you sure adding a 4th stick affects the performance of all memory, or is it just the 4th stick that's dragging the overall score down.

    My understanding is that regardless of the presence of the fourth stick, the first three will run in interleaved mode (stripping data across the 3 channels). The 4th stick obviously has no counterparts to interleave data across so it will run in SINGLE channel mode. Thus a memory benchmark that actually utilizes all available memory, will suffer, and yes, the fourth stick is dragging the benchmark down, but that doesn't necessary imply that the first 6GB are not fully interleaved and running at maximum.

    Just a suggestion since I don't know how Apple implemented things nor do I know how your benchmark works.
     
  15. rylin macrumors 6502

    Joined:
    Aug 18, 2006
    #15
    AFAIK, X58 will run memory in Triple channel or Dual channel.
    In other words, the fourth stick makes sure things get divvied up into pairs.
     
  16. Jouhne macrumors member

    Joined:
    Mar 6, 2009
    #16
    From what I remember the 3 RAM slots work in triple channel and the 4th in single channel... and the speed drop only when the 4th slot is working
     
  17. TrapOx macrumors 6502

    Joined:
    Dec 4, 2008
    Location:
    Denver
    #17
    Thats not possible. The memory controller only has 3 channels available. Where would it get the 4th channel and then why wouldn't it be able to run quad channel memory?
     
  18. IainH macrumors member

    Joined:
    Mar 11, 2009
    #18
    I dont suppose you have access to Shake or Nuke to test core and memory usage do you?
     

Share This Page