1. Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

i7 and multi-GPU scaling.

Discussion in 'Graphics Cards' started by OzzmanFloyd120, Dec 10, 2008.

?

Why does the i7 allow such great multi-GPU scaling?

Poll closed Dec 20, 2008.
  1. It's because the i7 architecture is superior.

    9 vote(s)
    32.1%
  2. The architecture of the chipset is superior.

    8 vote(s)
    28.6%
  3. Who cares, as long as it works!

    8 vote(s)
    28.6%
  4. I think it's from some other reason (post thoughts and reasoning)

    3 vote(s)
    10.7%
  1. OzzmanFloyd120

    OzzmanFloyd120

    Joined:
    Apr 3, 2008
    Messages:
    3,047 (1.31/day)
    Thanks Received:
    297
    Location:
    Pontaic Michigan
    It's no argument that since the introduction of the i7 core CPUs that there's been a leap in scalability of having 3-4 GPUs. This makes me curious though as to why the GPUs scale so well. It's been suggested that it's the memory bandwidth, and it's also been suggested that it's having so many threads, but I find this hard to believe because the former could be acheived by OCing the AMD64 line of CPUs and the latter never revealed any results from moving to one to two to four threads.
    My belief is that the scaling power doesn't lay in the CPU itself or even the memory controller, I believe that it has far more to do with the x58 chipset's architecture.
    I just want to see what you guy think in this matter. Maybe we all might learn something.
  2. Binge

    Binge Overclocking Surrealism

    Joined:
    Sep 15, 2008
    Messages:
    6,982 (3.22/day)
    Thanks Received:
    1,751
    Location:
    PA, USA
    The architecture of the chipset is pretty much dedicated to the PCI-E/peripherals instead of dealing with memory interface and PCI-E/peripherals.

    Agreed. It's the chipset architecture.
  3. trickson

    trickson OH, I have such a headache

    Joined:
    Dec 5, 2004
    Messages:
    6,494 (1.83/day)
    Thanks Received:
    956
    Location:
    Planet Earth.
    It's the architecture .
  4. OzzmanFloyd120

    OzzmanFloyd120

    Joined:
    Apr 3, 2008
    Messages:
    3,047 (1.31/day)
    Thanks Received:
    297
    Location:
    Pontaic Michigan
    Architecture... of which? The CPU or the chipset?
  5. trickson

    trickson OH, I have such a headache

    Joined:
    Dec 5, 2004
    Messages:
    6,494 (1.83/day)
    Thanks Received:
    956
    Location:
    Planet Earth.
    The CPU and the Chipset . Both !
  6. Solaris17

    Solaris17 Creator Solaris Utility DVD

    Joined:
    Aug 16, 2005
    Messages:
    17,134 (5.20/day)
    Thanks Received:
    3,544
    Location:
    Florida
    banking on chipset arc...which is why i cant wait to complete my build....mobo and ram needed. 2 GX2's in quad sli....sweeeeeeet.
  7. CrAsHnBuRnXp

    CrAsHnBuRnXp

    Joined:
    Oct 19, 2007
    Messages:
    5,466 (2.19/day)
    Thanks Received:
    636
    Just so everyone knows, I was the one that voted "who cares, as long as it works." :)
  8. farlex85

    farlex85 New Member

    Joined:
    Mar 29, 2007
    Messages:
    4,830 (1.79/day)
    Thanks Received:
    638
    I would imagine QPI's added bandwidth perhaps gives it an edge over fsb? So it would be both a and b, actually b is really a, unless you are referring to the x58 in particular.
  9. OzzmanFloyd120

    OzzmanFloyd120

    Joined:
    Apr 3, 2008
    Messages:
    3,047 (1.31/day)
    Thanks Received:
    297
    Location:
    Pontaic Michigan
    Yes, I'm referring to the x58 specifically.
  10. James1991

    James1991 New Member

    Joined:
    Jun 28, 2007
    Messages:
    267 (0.10/day)
    Thanks Received:
    38
    Location:
    Port Lincoln, South Australia
    i'm the second:D
  11. farlex85

    farlex85 New Member

    Joined:
    Mar 29, 2007
    Messages:
    4,830 (1.79/day)
    Thanks Received:
    638
    Ah, well then I voted wrong sorry. :(
  12. CrAsHnBuRnXp

    CrAsHnBuRnXp

    Joined:
    Oct 19, 2007
    Messages:
    5,466 (2.19/day)
    Thanks Received:
    636
    Im the type of person that doesnt care how it works I just want it to the way its supposed to. Once I set soemthing up, I could care less from there. I did my thing. Let it do its.
  13. James1991

    James1991 New Member

    Joined:
    Jun 28, 2007
    Messages:
    267 (0.10/day)
    Thanks Received:
    38
    Location:
    Port Lincoln, South Australia
    :toast:
  14. X2-3800 New Member

    Joined:
    Oct 21, 2008
    Messages:
    18 (0.01/day)
    Thanks Received:
    0
    I7 is as fast as the old Q9xxx if you test it at the same clock, Turbomode off(overclockingCheat), and Hyperthreading off. Hyperthreading is useless if you have software that don't use it.
    QPI and the integrated Memorycontroller don't bring much more power, because of the big caches.
    Many sources said that in the most games the new I7 brings not much more power.
    (cause of the lag of multicore and hyperthreading support)
    The I7 is overkill for people wich only playing games or doing office work.

    "My belief is that the scaling power doesn't lay in the CPU itself or even the memory controller, I believe that it has far more to do with the x58 chipset's architecture."

    I think so, too. The only real reason for the GPU scaling can be the chipset and its PCIeX performance.
  15. [I.R.A]_FBi

    [I.R.A]_FBi New Member

    Joined:
    May 19, 2007
    Messages:
    7,664 (2.89/day)
    Thanks Received:
    540
    Location:
    c:\programs\kitteh.exe

    i5 is just right then?
  16. X2-3800 New Member

    Joined:
    Oct 21, 2008
    Messages:
    18 (0.01/day)
    Thanks Received:
    0
    http://www.hardware-infos.com/news.php?news=2571
    Sorry for german laguage. They say that the massiv memoryperformance is at most useless.

    I think I5 is the better choice, it have only two memorychannels and is for sure less expensive than the I7. That should drop the price for the mainboards, too. Currently you have to pay >230€ till over 300€ for one I7 mainboard!

    With I5:
    1. You need no expensive 3 channel memorykits.
    2. Mainboards are much less expensive because of
    "Sockel mit 1156 statt 1366 Kontaktflächen " > Sockets with less Pins >less costs at
    motherboard design.
    3. less memorychannels > less power consumption

    The draw back is the decreased memoryperformance, but the DDR3 memory can only get faster with the time. Timings are going down and frequencys going higher.
  17. rizla1

    Joined:
    Oct 23, 2008
    Messages:
    482 (0.23/day)
    Thanks Received:
    9
    Location:
    ireland
    ddr3 hardly consumes much powerdoes it, id be happy with i5 still got 8 treads . i do 3d work photoshopings and heavy multitasking so would benifit from it even and just overclock it.
  18. DarkMatter New Member

    Joined:
    Oct 5, 2007
    Messages:
    1,714 (0.68/day)
    Thanks Received:
    184
    I think it's primarily the fact that they moved the memory controler to the CPU (that also made A64 fly) + more and far better caching system + improved latencies + better prefetching + better thread hadling + ...

    i7 improves so many things, it can't be just because of one. In one article I read in Real World Technology about CUDA but that explained GT200 too, they said that current generation of cards didn't perform as they should (2x the previous one) because the CPU couldn't generate all the required threads without introducin lag.

    Also I strongly believe in the possibility that PCIe 2.0 was improved in the X58. Just because a chip follows one standard it doesn't mean it can deal with it 100% effeciently. Just as with memory, you can use PC11200 memory but that doesn't mean the chipset will deliver that.
  19. lemonadesoda

    lemonadesoda

    Joined:
    Aug 30, 2006
    Messages:
    6,247 (2.14/day)
    Thanks Received:
    963
    I think there is a combination of:

    1./ Improved drivers of the GPU
    2./ Ability for QPI to allow competing demands on system memory (both CPU, GPU1 and GPU2) to be simultaneously serviced. Remember that if you have 2 GPUs you need to load ONE asset TWICE. Therefore 2GPUs have double the memory bandwidth demand. GPU2 cant draw until BOTH GPU1 AND GPU2 have loaded the asset.

    I think i5 might not scale as well with 2 GPUs due to point 2.

    A larger GPU memory (more cached onboard and less use of system memory) would reduce the effect of 2.

    Therefore 1GB or 2GB boards would scale better than 256MB or 512MB.

    Somebody with 2x Quadro FX4800/5600's, an i5 and an i7 can prove this theory. :toast:

    ANSWER: DDR3 + chipset + QPI architecture

    Where is my option in the poll? Doh!
  20. The Haunted

    Joined:
    Aug 5, 2007
    Messages:
    311 (0.12/day)
    Thanks Received:
    36
  21. kid41212003

    kid41212003

    Joined:
    Jul 2, 2008
    Messages:
    3,584 (1.60/day)
    Thanks Received:
    533
    Location:
    California
    It's because even the Yorkfield couldn't draw ALL the powers of 3 of 4 GPUs.

    More GPUs require more CPU powers.

Currently Active Users Viewing This Thread: 1 (0 members and 1 guest)

Share This Page