Does High-Speed DDR3 Help AMD's FX? Four 8 GB Kits, Reviewed

DDR3-1866 Performance

Retaining the reference DDR3-1600 bandwidth numbers, we can see how bumping each of the competitive module sets to DDR3-1866 boosts Sandra’s Memory Bandwidth benchmark. G.Skill's modules lead at this point, though by an insignificant amount.

Lower primary timings continue to push Team Group’s Xtreem DDR3-2800 to the top of our gaming chart, though Metro 2033’s memory needs appear satiated by any DDR3-1866.

G.Skill appears at the top of our DDR3-1866 application chart, but only because identically-performing products are arranged in alphabetical order. In reality the top three contenders yield the same average performance between the two tests, and have yet to prove noticeably advantageous against the reference DDR3-1600 CAS 9.

Create a new thread in the US Reviews comments forum about this subject
This thread is closed for comments
65 comments
    Your comment
    Top Comments
  • Great article this proves a lot of rumors!

    People can finally stop saying how unfair it is to test Amd with slower ram when it supports higher speed ram even though higher speed ram cost more money. It would actually be unfair to Intel its not their fault Amd needs higher speeds to compete(which it doesn't) with Intel's 1333 ram controller.

    Ram speeds do how ever make a big difference on APU's.
    16
  • noob2222take a small enough sample to make sure the results don't prove a thing. why not use some programs that do respond to memory speed?
    The test here was set up long ago to use 50% GPU-limited and 50% platform-limited games. Replace GPU with CPU for the other tests, and the hope was to find 50% of the benchmarks with noticeable performance gains. Of course that's based on a real-world scenario, where both GPU and platform-limited games coexist.

    Notice in the chart you showed, the difference between 2133 and 1866 is rather small. The baseline in this test was DDR3-1600 CAS 9 because that's the slowest memory a performance builder would use. In other words, the 1333 and 1066 results are irrelevant.

    The benefit you're seeing in some articles is reduced real-time latency, which I'm fairly certain is discussed in this article. DDR3-1333 CAS 7 has the same real-time latency as DDR3-2666 CAS 14, since the cycle time is inverse of frequency.

    If we assume for only a moment that DDR3-1600 is barely fast enough to fill the bandwidth needs, most of the difference over 1600 actually measures how long the CPU is waiting for instructions. Now I wouldn't actually assume 1600 is enough, I only assume that 1600 is the minimum speed that goes into most performance builds. After all, who would buy 1333 to save money in a performance machine when 1600 is so cheap?
    14
  • would love to see the same review for Ivy Bridge processors with and without iGPU HD4000.

    and do include more tests in the review.
    11
  • Other Comments
  • Interesting!. Still would wait for Pile driver. But since i do have a Kingston 8GB kit and a Sabertooth 990FX... may just have to try it one day
    2
  • Great article this proves a lot of rumors!

    People can finally stop saying how unfair it is to test Amd with slower ram when it supports higher speed ram even though higher speed ram cost more money. It would actually be unfair to Intel its not their fault Amd needs higher speeds to compete(which it doesn't) with Intel's 1333 ram controller.

    Ram speeds do how ever make a big difference on APU's.
    16
  • would love to see the same review for Ivy Bridge processors with and without iGPU HD4000.

    and do include more tests in the review.
    11
  • More of the cheap stuff is my mantra.
    7
  • sarinaideMore of the cheap stuff is my mantra.
    Yeh yeh, but I really can't complain about paying $70 for 8GB of DDR3-2133, especially when it overclocks like that.
    8
  • AMD has enjoyed the performance benefits of an integrated memory controller for more than twice as long as Intel. And yet, it seems that Intel sets today's standard for acceptability.

    Replace the last word with "performance" crash.

    Good article which gets answers a couple of questions people have been asking - well done !!

    :)
    3
  • Thomas thanks for the review.

    First note:
    *KHX2400C11D3K4/8GX - is a Quad Channel kit specifically designed for the X79 ; link http://www.kingston.com/us/company/press?pagename=n1111c&year=2011&prLanguage=english_emea
    *WQ213UB4G vs WQ213UX16G (kit) - is ONE stick of RAM (not a matched kit) and its 'kit version' is Quad Channel kit specifically designed for the X79 ; link http://www.supertalent.com/datasheets/WQ213UX16G.pdf

    After that it makes sense that those (2) "kits" (in one case 4 individual sticks) failed @ Rated, and I wasn't interested in researching the others. I've seen other more in-depth testing on the AMD FX-8150 indicating that overall there's a slight advantage for the DDR3-1866 kits on a most of the AMD FX processors. Though @ 4AM 'to me' I'll edit this post after a few cups of coffee.

    The 'ideal' AMD kits have JEDEC 'SPD' @ Rated 'cloned' (XMP) encoding which is optimized for AMD and sometimes slightly different so you're not 'stuck' as you put it @ DDR3-1333. Those are the better KITS to test on the AMD FX lines.
    3
  • Ram sppeds on a Bulldozer CPU means little and this just proves that so it doesn't matter if you are using DDR3 1600 ram or buying high performance RAM that can overclock to DDR3 2800 your not really seeing hardly any improvement with Bulldozer by doing so.
    7
  • Good article, very informative!
    Thanks :)
    3
  • RAM speed also has a HUGE impact in controlled cache environments:
    http://thessdreview.com/our-reviews/romex-fancycache-review-ssd-performance-at-13gbs-and-765000-iops-in-60-seconds-flat/

    Even the best SSDs can barely muster 100,000 IOPs, yet even slow RAM can easily exceed 700,000 IOPs.
    Bandwidth, throughput, and latency are equally insane. RAM puts SSDs to shame exponentially more than SSDs put hard drives to shame. Contrary to what many "enthusiasts" believe, RAM timings are virtually irrelevant. Frequency is by far, much more important.
    0
  • jaquithThomas thanks for the review.First note:*KHX2400C11D3K4/8GX - is a Quad Channel kit specifically designed for the X79 ; link http://www.kingston.com/us/company [...] glish_emea*WQ213UB4G vs WQ213UX16G (kit) - is ONE stick of RAM (not a matched kit) and its 'kit version' is Quad Channel kit specifically designed for the X79 ; link http://www.supertalent.com/datasheets/WQ213UX16G.pdfAfter that it makes sense that those (2) "kits" (in one case 4 individual sticks) failed @ Rated, and I wasn't interested in researching the others. I've seen other more in-depth testing on the AMD FX-8150 indicating that overall there's a slight advantage for the DDR3-1866 kits on a most of the AMD FX processors. Though @ 4AM 'to me' I'll edit this post after a few cups of coffee.The 'ideal' AMD kits have JEDEC 'SPD' @ Rated 'cloned' (XMP) encoding which is optimized for AMD and sometimes slightly different so you're not 'stuck' as you put it @ DDR3-1333. Those are the better KITS to test on the AMD FX lines.
    Read page 2. They're all Intel-optimized kits because that's the only thing any of these companies have. If you read something about an AMD-optimized kit being for sale somewhere, it's probably either a niche manufacturer or simply a marketing lie.
    9
  • If you're going to spend extra money on memory, why not use the extra money for a real CPU? Meaning, from Intel, AMD doesn't make anything remotely competitive, and if you're getting performance parts to try to make it perform, you're probably better off starting with a processor that performs better in the first (second, and third) place.

    It's like putting a diamond in an aluminum ring. It makes no sense.
    -6
  • fx needs ddr3 2133 kits to reach 20~ gb/s on sandra bench with while a sandy bridge pentium can reach the same bw @ddr3 1333.
    http://media.bestofmicro.com/G/Y/327202/original/sandra%20mem.png
    amd apus are different story. those perform better in games with faster memory.
    only posted in case some wonder how much intel gets out of rams.
    7
  • great article, 2 games and 2 programs that don't respont to memory speed must represent all games ever made.

    This article proves one thing, not every program responts to memory speed.

    http://vr-zone.com/articles/amd-fx-8150-memory-scaling-investigation--feeding-the-bulldozer/13704.html
    http://www.madshrimps.be/articles/article/1000220/AMD-FX-8150-Bulldozer-CPU-Review/4#axzz1wX1gAnfF

    Ill match your 4 cherry picked benchmarks against 2 websites that state otherwise.

    Throw civ V into the mix and your way off.
    -4
  • noob2222great article, 2 games and 2 programs that don't respont to memory speed must represent all games ever made. This article proves one thing, not every program responts to memory speed.http://vr-zone.com/articles/amd-fx [...] 13704.htmlhttp://www.madshrimps.be/articles/ [...] z1wX1gAnfFIll match your 4 cherry picked benchmarks against 2 websites that state otherwise.Throw civ V into the mix and your way off.
    Cherry picked for what, to prove that faster memory helps? Well guess what, it didn't help much. Or did you mean cherry picked for this article? Are you making blind accusations without even looking to see which benchmarks were used in the last several memory articles?

    Please, at least try to be honest. You already knew the Madshrimps article was explicitly dishonest because it used CAS 8 timings at all speeds (even 2133). Real-world memory supports tighter timings at lower frequencies, and requires looser timings at higher speeds. Madshrimps gimped the low-speed tests and boosted the high speed configuration intentionally.
    10
  • take a small enough sample to make sure the results don't prove a thing. why not use some programs that do respond to memory speed?


    6
  • noob2222take a small enough sample to make sure the results don't prove a thing. why not use some programs that do respond to memory speed?
    The test here was set up long ago to use 50% GPU-limited and 50% platform-limited games. Replace GPU with CPU for the other tests, and the hope was to find 50% of the benchmarks with noticeable performance gains. Of course that's based on a real-world scenario, where both GPU and platform-limited games coexist.

    Notice in the chart you showed, the difference between 2133 and 1866 is rather small. The baseline in this test was DDR3-1600 CAS 9 because that's the slowest memory a performance builder would use. In other words, the 1333 and 1066 results are irrelevant.

    The benefit you're seeing in some articles is reduced real-time latency, which I'm fairly certain is discussed in this article. DDR3-1333 CAS 7 has the same real-time latency as DDR3-2666 CAS 14, since the cycle time is inverse of frequency.

    If we assume for only a moment that DDR3-1600 is barely fast enough to fill the bandwidth needs, most of the difference over 1600 actually measures how long the CPU is waiting for instructions. Now I wouldn't actually assume 1600 is enough, I only assume that 1600 is the minimum speed that goes into most performance builds. After all, who would buy 1333 to save money in a performance machine when 1600 is so cheap?
    14
  • CrashmanThe test here was set up long ago to use 50% GPU-limited and 50% platform-limited games. Replace GPU with CPU for the other tests, and the hope was to find 50% of the benchmarks with noticeable performance gains.


    the problem is both of those games, dirt 3 and metro 2033 are gpu limited. http://www.techspot.com/review/403-dirt-3-performance/page7.html
    -6
  • noob2222the problem is both of those games, dirt 3 and metro 2033 are gpu limited. http://www.techspot.com/review/403 [...] page7.html
    GPU limited at max quality and 1920. This test was at medium resolutions and details (the "high" setting in these games is actually mid-high). DiRT 3 in particular is supposed to be GPU-bottlenecked only at its highest settings.
    10
  • If you have a copy of civ V, try it, you will laugh at the results, 1333 to 2133 is nearly 25% faster on the full render. There are benefits to faster memory, limiting the benchmarks limits the results.
    0