Sign in with
Sign up | Sign in

Core i7 Memory Scaling: From DDR3-800 to DDR3-1600

Core i7 Memory Scaling: From DDR3-800 to DDR3-1600
By

According to Intel, the Core i7 processor only officially supports DDR3-1066 memory, but in fact, the family is capable of running memory speeds from DDR3-800 to DDR3-1600. Even more clock speed (to the tune of 2,133 MT/s, as seen from our Core i7-975 review) is possible through overclocking and with additional voltage, which Intel advises against to avoid damage to the Core i7’s memory controller. But which speed is best? Does it make sense to invest in expensive, high-speed RAM?

While we ran some memory scaling numbers in the i7-975 story, we only had time to test bandwidth and one real-world data-intensive metric (MainConcept). Interested in how Intel's fastest CPU benefits from the balance between throughput and latency, we're running a more exhaustive benchmark suite on a number of different memory configurations for more insight.

Single-, Dual-, Triple-Channel

Memory bandwidth has always been an important determinant of overall CPU performance. DDR3 and DDR2, which represent the third and second generation of double data rate (DDR) memory, are most commonly used today. Every new DDR memory generation has been capable of eventually running at faster clock speeds (requiring new memory modules and platform support) than its predecessors, but it is also possible to increase bandwidth by widening the data path. Therefore, dual-channel memory operation was introduced by AMD and Intel in 2002, doubling the data path from 64 to 128 bits. Intel took the next step in late 2008, when triple-channel DDR3 was introduced with the launch of the Core i7 processor to the desktop market.

Cache Versus Memory Performance

However, Core i7 (and equivalent Xeons) will remain the only triple-channel products in Intel’s portfolio. Upcoming platforms and processors will stick with two channels, and they do so for a reason. While dual-channel memory made quite a difference a few years ago, increasing cache capacities helps soften the impact of insufficient memory bandwidth. This means that the difference among triple-, dual-, or even single-channel memory isn’t as big as it was several years ago.

Triple Channel, Sure. But Which Speed and Timings?

There is no necessity to discuss memory capacity and configuration thanks to lower DDR3 prices. For Core i7, it has to be triple-channel memory, ideally a 3 x 2 GB DDR3 memory kit—these 6 GB kits currently offer the best bang for the enthusiast buck.

We already performed a thorough analysis on the performance and behavioral differences of 3 GB vs. 6 GB vs. 12 GB RAM on a Core i7 system. Now it’s time to look at popular RAM speeds and timings using the most popular 6 GB RAM capacity. Which DDR3 memory configuration is best?

Display 35 Comments.
This thread is closed for comments
  • -2 Hide
    Tindytim , June 9, 2009 6:13 AM
    Quote:
    We already performed a thorough analysis on the performance and behavioral differences of 3 GB vs. 6 GB vs. 12 GB RAM on a Core i7 system.

    I disagree.

    You only used apps that the average user would. 12GB of RAM is more useful to professional applications.

    And you didn't even turn paging off.
  • 6 Hide
    Why_Me , June 9, 2009 7:13 AM
    Great review...and for all those suckers who spent god only knows how much money on that over priced Corsair Dominator DDR3...ha ha ha ha...SUCKERS!
  • 2 Hide
    doomtomb , June 9, 2009 7:27 AM
    TindytimI disagree.You only used apps that the average user would. 12GB of RAM is more useful to professional applications.And you didn't even turn paging off.

    I think everybody knows that more RAM would be useful for "professional applications" where there are endless possibilities to eat up your memory. Everything from virtualizing OSes to medical imaging.
  • 3 Hide
    apache_lives , June 9, 2009 7:30 AM
    TindytimI disagree.You only used apps that the average user would. 12GB of RAM is more useful to professional applications.And you didn't even turn paging off.


    Also benchmarking one app at a time is utterly useless, same goes for processor benchmarks - thrash that memory subsystem and rattle that architecture to get a true result - that low latency memory may just pay off when multiple cores are thrashing the ram/fsb etc.
  • 0 Hide
    Tindytim , June 9, 2009 7:31 AM
    doomtombI think everybody knows that more RAM would be useful for "professional applications" where there are endless possibilities to eat up your memory. Everything from virtualizing OSes to medical imaging.

    Of course it would be, but the question is by how much? I do 3D work with my Desktop, and I get some lag when working with large and/or complex scenes. It would have been nice if they had used a few applications that made used of that RAM, and would be used by Desktop users (Blender anyone?).
  • 6 Hide
    neiroatopelcc , June 9, 2009 8:16 AM
    These results look very little different from the amd results. Guess it's as it always has been. Get the amount of memory you need, and the frequencies you need for whatever oc you aim at (if any).
  • 0 Hide
    Summer Leigh Castle , June 9, 2009 8:38 AM
    Good stuff. I find these consumer focused articles a great read. Just like the AMD article, there appears to be minimal performance gains in real world applications with the faster (and usually more expensive) DDR3 RAM. Looks like DDR3-1333 and low latency is the sweet spot for price-performance conscious shoppers. Thanks!
  • -2 Hide
    anamaniac , June 9, 2009 9:00 AM
    DDR3-1600 is cheap.
    I have no reason not to be interested in it.

    I'm sure you're not going to upgrade your memory too soon, so might as well. It's not overly expensive, so I doubt you'll regret it.
    Hell, at $10/gig for ddr2 800, and $15/gig for ddr3 1600... I think it's easy enough.

    Nice article on the benefits of DDR3 at different clocks. Thanks.
  • -2 Hide
    bin1127 , June 9, 2009 11:24 AM
    good article. as with all computer components, most users just need to find the best price/performance parts and they'll have themselves a great running computer. there's no need even for enthusiast gamers to spend 3x the money necessary on some miracle ram. even if they gave 10 more frames on every game, you can easily buy another videocard for the extra amount.
  • 0 Hide
    alert101 , June 9, 2009 11:45 AM
    Does the QPI link affect memory performance? If it does, then I would have liked to see the i7 920 (which has a lower QPI speed) results in this same benchmark suite.
  • 2 Hide
    coolkev99 , June 9, 2009 12:51 PM
    Great article. This is the kind of shit people want to know.
  • 2 Hide
    burnley14 , June 9, 2009 1:22 PM
    This is probably the single most relevant article you could have written for me right now. Thanks Tom's!
  • 3 Hide
    NocturnalOne , June 9, 2009 2:14 PM
    Good article. Especially the day *after* I just got my i7 rig at newegg :)  I picked up 2x3 GB of OCZ Platinum 1600 7-7-7-24. From this article I could have saved a few bucks with cheaper memory but at $80 after MIR and free shipping it didn't break the bank. My primary application is astronomical image processing. Very memory and CPU intensive. Actually it's disk intensive as well. Intensive everything I guess! :) 
  • 0 Hide
    mapesdhs , June 9, 2009 2:41 PM

    NocturnalOne, if your app is disk intensive aswell, try looking into picking up
    an U320 RAID card and some 2nd-hand 15000rpm SCSI disks. You'd be amazed what's out
    there. I managed to get an unused 300GB 15K Fujitsu drive for just under $120,
    though it's easier to obtain 146GB 15Ks at good prices.

    Access times of this type of drive knock the socks off SATA, and there's usually
    lots of RAID cards on eBay; I use both PCIX and PCIe cards. For raw storage, I just
    have two 1TB SATAs (Samsung HD103UJ) in different systems; this drive has a decent
    sustained I/O, but an average access time 3X slower than 15K SCSI.

    A SCSI setup should give your work a hefty speed boost if its I/O-bound. SCSI is
    far too expensive new, but 2nd-hand it can be a real bargain.

    My main experimental setup (Tezro) has 8 x 146GB 15K, which gives 540MB/sec random
    write speed, 400MB/sec random read, 650MB/sec sequential write, 480MB/sec sequential
    read. The cards I normally hunt for are the LSI 22320-R PCIX and the LSI 22320IE
    PCIe, though the Tezro has several QLA12160s instead just now.

    My PC has an LSI 22320-R with 12 x 146GB 15K Maxtor Atlas 15K II (external unit),
    while the system disk (Seagate 146GB 15K ST3146855LC) is connected to an LSI 20320IE
    PCIe card. I'm going to build an i7 system in Jul/Aug, for which I have an LSI
    22320IE PCIe card ready & waiting.

    Are you dealing with very large 2D images by any chance?

    Ian.

    PS. If your mbd doesn't have PCIX slots, then just hunt for PCIe cards instead;
    they're slightly faster anyway.

  • 0 Hide
    NocturnalOne , June 9, 2009 3:01 PM
    Hey Ian,

    thanks for your reply. I was actually considering getting an SSD but decided to hold of until prices drop a little. Then I'll decide if I'll add one as a data drive or move the OS to it.

    The disk intensive part is what's called calibration and stacking. Dozens (sometimes hundreds) of images are averaged to increase the signal to noise ratio. This is also quite CPU intensive and makes full use of every core so there was no doubt I'd benefit from an i7. Each image is 11MB (3000x2000 points). Once I get the new system put together I'll figure out where the bottleneck is and then act appropriately. Getting a SCSI array is a good idea. I picked the MSI X58M motherboard as I wanted a small, relatively quiet package with the Antec mini P180 case.

    Post processing of the images (which are now 3000x2000x3x32bit or about 60MB in TIFF format) is mostly CPU bound although writing and reading swapfiles takes some time too. It's already pretty quick on my current system so it'll probably be fine with the 1TB 7200.12 I picked. Again all cores are used to do complex transformations on the images so an i7 was the natural choice.

    This stuff is just for my hobby/obsession otherwise I'd plop down even more cash :) 
  • -2 Hide
    Kill@dor , June 9, 2009 3:48 PM
    In real world apps you won't see a difference, but in gaming you will with higher CPU clocks and Memory OC...

    If you have a board that has a QPI mulitplier of x44, 250 BCLK, CPU Muliplier 16 and Memory multiplier/divider of 8 or 15 that should hit DDR3-2000, with 4GHz OC. Your QPI should or would run anywhere between 9GT'z to 10GT's effective, depending on your voltages, and the core 920's only need 1.27v or so to maintain speed. The QPI voltage is the tricky part because your system becomes unstable as soon as you increase the QPI multiplier and voltage needs to go up on that to maintain stability.

    Perhaps buying DDR3-1600 or less is better, but if you are gaming with intense graphic games, you need a good GPU and good RAM.

    Just my 2 cents...
  • 0 Hide
    cjl , June 9, 2009 3:51 PM
    Why on earth would you need a 10GT/s QPI?
  • -1 Hide
    chripuck , June 9, 2009 5:06 PM
    kami3kI got the fastest memory they listed at 6GB for 90 dollars, yea Why_Me I'm such a sucker. Idiots, gotta love them.


    Yeah no joke, I paid $95 three months ago for 6 Gigs of DDR3, 8/8/8/21. I could have saved a whopping $15 at most by going with "slower" RAM. I'm glad to pay an extra $15 for even a CHANCE that I might get that 7% boost seen in LFD.
  • 5 Hide
    Why_Me , June 9, 2009 5:12 PM
    6 gigs of Dominator starts at $160 - $300+ at newegg. That's what I was referring to. Nobody has to be a rocket scientist to figure that one out. I paid $90 for my G.SKILL 1600.
  • -4 Hide
    DjEaZy , June 9, 2009 7:13 PM
    ... it's funny...
Display more comments