X58 Triple Channel DDR3 Memory Roundup! 8 Mid-range Kits Tested

Memory by thorgal @ 2009-04-09

In the ever raging sales battle among memory manufacturers, the Core i7 launch must have been the wind that shook the barley. While everyone was convinced i7 would be the real start of DDR3 technology, it also was a ´back to the drawing board´ call for most, having to come up with lower voltage memory parts - with decent performance - in a hurry. Four months into the launch now, we present you with a roundup of 8 mid-range kits from 8 different manufacturers. Begun, the new battle has.

  • prev
  • next

Test methodology : a new ballgame

A new ballgame

A new ballgame it is, so what will we be testing, and how will we go to work ? Several issues had to be resolved : which components to choose ? Which speeds to test ? At what Latencies ? How will we overclock the memory ? Etcetera. This is the page where I’ll try to explain the whats and whys of our test methodology.

Overclocking the Core i7

I’m not going to delve into an overclocking “howto” on this page, this goes way beyond the scope of the review. This review does require some basic knowledge about the i7 architecture and overclocking (dis)abilities however, so if you’re new to i7, I’d gladly point you to the following reviews :

  • Our own launch review
  • Massman’s excellent X58 motherboard roundup
  • Anandtech’s ”Core i7 launch review”, ”Nehalem: The unwritten chapters” and ”A tale of 2 processors”

    Some important facts to remember when overclocking :
  • front-side-bus is dead – hurray for base clock, we call it bclk from here on. Increase it to increase the cpu speed, nothing new on that front.
  • quick path interconnect or QPI is the interface that is used by the processor to talk to the main chipset – the X58. We don’t talk about the northbridge anymore since… well… front side bus is dead remember.
  • the uncore : to make it easy : this is the part of the cpu that isn’t the core. The uncore consists out of “the rest of the cpu” : the level 3 cache, the qpi link and the memory controller, which are all on chip. With the uncore comes an uncore clock speed, which is, depending on the cpu model, running at 2.13Ghz or 2.66Ghz out of the box.
  • memory dividers : like core 2, or older cpu’s depending on front-side-bus, the memory runs at a speed that is a multiple of the base clock. This multiple is the divider, and this is at least 3 (called 2:6) and at most 8 (called 2:16). This results in memory speeds at 133bclk of 800Mhz up to 2000Mhz DDR
  • the core clock speed, the qpi link and the uncore clock are always a multiple of the base clock
  • the uncore clock has to run at a multiplier least twice that of the memory. If we apply a memory multiplier of 2:12, the uncore has to run at least at 24x bclk. This is very important for this particular review. More on this later by the way.

    The CPU

    As you might have guessed by now, more than ever before the CPU is the center of PC-setup. As the memory controller is now integrated, it is actually the CPU that “decides” how the memory will react, to voltage as well as to latencies and speeds (and overclocks). This cannot be stressed enough: a bad CPU can really spoil the fun here, and limit your headroom drastically.

    As this is a Core i7 roundup, at the time of writing we had the choice between 3 cpu’s : the i920, i940 and extreme edition i965. For this review we chose the i965, not because we like to spend money, but because the i965 has a couple of advantages :
    - unlocked multipliers : this one’s obvious : as the multiplier is unlocked, we can choose multi’s higher than 22 (the limit of the 940). We’ll be using the 20 and 24 multiplier in this review.
    - A higher QPI speed (6,4Gb/s vs. 4,8Gb/s) and the advantages that come with it : it’s not the base QPI speeds that are of interest, but rather the increased potential that resides inside the CPU that interest us. This together with the…
    - higher uncore clock (2.66Ghz vs. 2.13Ghz). As we mentioned it’s the CPU that decides how well memory will clock at a given voltage, and the i965 should have (and has as far as I’ve experienced) more headroom on the QPI and uncore side compared to the i920 and i940. Result : more memory overclock headroom at lower voltages.
    - PS: unlocked memory ratio’s : originally Intel planned to lock the memory ratio’s of the i920 an i940, and provide a maximum of 2:8 (4x memory multi). Most of the ES samples, even those at launch, were built this way. A limit of 2:8 would’ve meant you needed 200bclk to reach 1600Mhz ram speeds. Fortunately, this does not seem the case anymore for retail 920/940 models.


    Finally, a note about the turbo setting. As we’ll be using different bclk speeds, we don’t want to use the turbo setting during our testing, as it would mean that the speeds under load would differ. We disabled the turbo setting in bios at all time, but we did include hyperthreading (HT) for all tests.

    The motherboard

    When I started the i7 memory roundup, the choice of motherboards was still very limited. All the motherboards are based around Intel’s X58 chipset at this time, and it doesn’t look like this will change in the next few months. All the motherboards also reach about the same overclock potential, not counting the Intel “smackover” than, because that one is a bit lacking in the OC department. All motherboards reach bclk limits between 200 and 222bclk at this time, not counting the few individuals that have been able to crack this border and went up to around 230. Bringing down the 222 border is actually very hard to do, you have to be lucky with your motherboard and cpu.

    This means that any X58 board will do, it all boils down to preference, included features, and, in my case, bios. I decided to go with Gigabyte for this review because the new bios these guys designed is very user-friendly, and still has all the options one could wish. The EX58-Extreme is their top end offering for the time being, and differs only from the UD5 in the chipset cooling design. Both are great choices.

    If you’re in doubt about which one to get, have another look at Massman’s i7 motherboard roundup featuring the Gigabyte as one of the contestants.

    …and the rest

    The rest of the setup is nothing special : we decided to use an 4870X2 as a graphics card for the review, and Tones thoughtfully provided us with a Gigabyte card.
    As a boot and test hard drive we used a standard Western Digital Raptor drive, an older generation drive actually : the WD800.
    Finally, for power we used the excellent “little” OCZ 1000W Proxstream, a PSU that hasn’t let me down in the two years that I’ve been using it. A little noisy perhaps, but just great for a high end test setup.

    Madshrimps (c)


    Test methodology

    Now that we know which hardware to test on, let’s talk about the settings we use. For our benchmark suite we selected two settings to compare our modules with. The choice of core clock, memory ratio’s, uncore and qpi clocks, and corresponding voltages makes the choice almost unlimited. Nevertheless, we went with the two following bios settings

    1. 133 bclk x 24 (3,2Ghz)

    For this test suite we decided to keep the bclk and cpu multiplier to stock speed., meaning 3.2Ghz. Turbo mode is disabled as discussed above.

    For the memory speed we could’ve run the memory at the 2:12 ratio meaning it would run at its rated speed of 1600Mhz. We decided not to do that, and reserve the second test for rated memory speeds. Instead we went for the 2:10 multiplier, resulting in a 1333 memory speed. As the memory ran a little slower we could tighten the latencies a bit, resulting in cas 7 or even cas 6 testing in most cases. In any case we let the memory kits run at their best possible timings. One other thing to note : we only used a command rate of 1T with the memory kits. This provides a little better performance out of the box, and is the “standard” setting that is used by the motherboard. Relaxing to command rate 2T will get you a little extra headroom or lower timings, we chose the 1T setting however and tested all memory kits in the same way.

    The voltages are the following (bios values, measured voltages are slightly lower) :
  • vcore : 1.175V
  • vQPI : 1.395V
  • vDimm : 1.66V

    The resulting speeds :
  • CPU : 3.2Ghz
  • QPI : 6.4 Gb/s
  • Uncore : 2.66Ghz
  • Memory : 1333Mhz

    A small note about the QPI voltage : the applied 1.395V is slightly exceeding Intel’s recommendations (1.35V max), this is due to the fact that I wanted to eliminate a possible QPI bottleneck when overclocking the memory (or when sharpening the timings).

    2. 200 bclk x 16 (3,2Ghz)

    In this test we try to mimic a core i920 overclock setting, as this cpu will no doubt be the most popular. The increased bclk results in a faster QPI speed, and because we kept the memory multiplier at 2:10 (resulting in exactly the rated speed of 1600Mhz this time), this resulted in a higher uncore speed (4Ghz). Again, each memory kit runs at its best possible timings.

    The voltages are the same as before (bios values, measured voltages are slightly lower) :
  • vcore : 1.175V
  • vQPI : 1.395V
  • vDimm : 1.66V

    The resulting speeds :
  • CPU : 3.2Ghz
  • QPI : 7.2Gb/s
  • Uncore : 4.00Ghz
  • Memory : 1600Mhz

    Madshrimps (c)


    Overclocking

    We wouldn’t be Madshrimps if we didn’t go on overclocking our guinea pigs. Like we did in the past, we try to give you an idea of the headroom at various cas levels : some memory likes low latencies better, others like high latencies at higher speeds. The following OC settings were used :

  • Maximum OC cas 7

  • vcore : 1.200V
  • CPU multi : 19-20
  • Memory divider : 2:10
  • Uncore multi : 20
  • QPI setting : x44
  • vQPI : 1.455V
  • vDimm : 1.66V
  • Memory timings : 7-7-7-21-1T


  • Maximum OC cas 8

  • vcore : 1.200V
  • CPU multi : 21
  • Memory divider : 2:12
  • Uncore multi : 24
  • QPI setting : x44
  • vQPI : 1.455V
  • vDimm : 1.66V
  • Memory timings : 8-8-8-24-1T


  • Maximum OC cas 9

  • vcore : 1.200V
  • CPU multi : 19-20
  • Memory divider : 2:12
  • Uncore multi : 24
  • QPI setting : x44
  • vQPI : 1.455V
  • vDimm : 1.66V
  • Memory timings : 9-9-9-27-1T


  • OK, now that the boring stuff is out of the way, let’s meet our test candidates....>
    • prev
    • next
    Comment from Kougar @ 2009/04/10
    Quote:
    I decided not to include the results in this review yet, as there wasn't actually anything wrong with the kits at their rated speeds. Both did the rated speeds and timings without any issues, they just didn't overclock to any satisfying level.
    This has been observed with Corsair Dominator kits in other reviews. Link

    I was just about to buy a kit of OCZ Platinum, nice to see the overclocking results here, thank you. Should probably mention the 6GB OCZ Platinum 1600MHz kit is available for $80 shipped AMIR, cheapest 1600MHz kit of all those tested.
    Comment from blind_ripper @ 2009/04/10
    80$ , thats cheap!

    btw u got a pic mixed up @ page 8 from the OCZ's, cas 7 has the cas8 pic .
    althoughe very nice work thorgal
    Comment from thorgal @ 2009/04/10
    Quote:
    Originally Posted by Kougar View Post
    This has been observed with Corsair Dominator kits in other reviews. Link

    I was just about to buy a kit of OCZ Platinum, nice to see the overclocking results here, thank you. Should probably mention the 6GB OCZ Platinum 1600MHz kit is available for $80 shipped AMIR, cheapest 1600MHz kit of all those tested.
    Yes, I read the xbit article, they come to about the same conclusion indeed.

    The Patriots can also be had for $79,99 now at Newegg, could be the OCZ's are equally cheap somewhere (they're $89,99 at Newegg).
    Comment from leeghoofd @ 2009/04/10
    Great work Ram "Torsten" Guy
    Comment from Kougar @ 2009/04/10
    Yes, but the Patriots are CAS 9. OCZ Platinum is CAS 7 (better than anything Corsair offers too).

    Didn't hear it from me: Newegg Promo code EMCLRLV42 for $10 off OCZ3P1600LV6GK ($80). Even less if buying NIB on ebay.
    Comment from Jaco @ 2009/04/11
    good article / review .

    Still reading ... lots of numbers

     

    reply