Memory Bandwidth Scaling

Memory bandwidth normally improves with increases in memory speed and reductions in memory timings. This is best illustrated using Performance Scaling charts.

Since results for high-end memory were so close at all tested speeds, the scale range was reduced to better show the small differences in these memory benchmark results. Please keep this in mind when viewing the charts, since a normal zero scale would make performance differences appear much smaller than these expanded scale charts. Test values for memory at each speed are included below each chart for reference.

To evaluate memory bandwidth SiSoft Sandra 2007 Professional was used to evaluate standard Buffered performance and Unbuffered memory bandwidth.

The most widely reported Sandra score is the Standard or Buffered memory score. This benchmark takes into account the buffering schemes like MMX, SSE, SSE2, SSE3, and other buffering tools that are used to improve memory performance. The Buffered benchmark does not correlate well with real performance differences in games, so our memory bandwidth tests have always included an Unbuffered Sandra memory score. The Unbuffered result turns off the buffering schemes, and we have found the difference in Unbuffered results correlate well with real-world performance differences.

In both Buffered and Unbuffered tests Corsair Dominator 10000 performance curves compare to the best tested to the top tested speed of DDR2-1066 on the ASUS 975X based motherboard. Dominator 10000 requires slightly slower timings at DDR2-1066 than the top OCC Flex and Dominator PC2-8888. This translates into bandwidth that is lower than the other top memories at this speed. While the differences are extremely small, and don't show well in the charts, both Standard and Unbuffered bandwidth is at the low end of tested memory at all speeds. This high-speed memory is binned for top speed and not low latency, and you can see that reflected in slightly lower bandwidth in test results.

Performance with Version 1.5 of Super Pi was also compared. The Super Pi test compared the time to calculate 2 million places of Pi at the different memory speeds. This benchmark measures pure number crunching.

Super Pi results continue to improve to the highest speed that could be reached with stability with Corsair Dominator 10000 and other tested memories. What also stands out is how close the results with all the memories really are from DDR2-400 to DDR2-1066.. This is mainly a reflection of the fact that the compared memories are based on the same base Micron memory chips.

Stock Memory Performance Game Performance Comparison


View All Comments

  • OBWan - Wednesday, November 14, 2007 - link

    It appears that the "highest memory speed" chart on page 7 is not an objective comparison. Besides the top 3 bars that were tested with the overclock-friendly 680i platform, all the rest were tested with old chipsets e.g. 975X. The old chipsets actually hindered the max overclocking potential of those other DDR2's while at the time of their reviews, the editors believed those DDR2's had reached their speed limit, which was an unsighted mistake. So please delete this biased comparison chart and re-do all the tests with the same 680i platform. The readers in this forum deserve not to be cheated. Reply
  • yacoub - Thursday, February 1, 2007 - link

    If those timings relax much more, they'll be asleep.
    5-5-5-18? hehe =)
  • phil verhey - Thursday, February 1, 2007 - link

    i had buffalo pc2-10000 3.5 months ago.. and i still overclocked it to 1300mhz at 5-5-5-18 .. talk about an either ignorant, or biased article. Reply
  • Wesley Fink - Friday, February 2, 2007 - link

    The Buffalo 10000 was announced in Japan a couple of months ago, and it was widely reported that DDR2-1200 and DDR2-1250 were "coming". We have reviewed Buffalo memory in the past and we have yet to receive Buffalo PC2-10000 samples for review. We also have not seen a single review of Buffalo PC2-10000. PC2-10000 is also not listed on the Buffalo website - the highest rated Firestix there are PC2-8500.

    The Corsair PC2-10000 is a production product with warranted PC2-10000 performance. We ALSO reached DDR2-1300 and DDR2-1315 several months ago, with OCZ and Corsair, but neither product wass rated or warranted at PC2-10000.

    I have no doubt the Buffalo PC2-10000 is likely a good product. It is just we have neither seen or received for testing Buffalo PC2-10000 - we have only seen press announcements.
  • VooDooAddict - Wednesday, January 31, 2007 - link

    I'd like you to also show a zero based benchmark graph. I understand that you mention it in the article ... but too many people will just look at the graphs. Budget focused people also should see how little going down to 667 or even 533 effects the performance (I try to steer people to 800 just for a little more future proofing of their memory investment).

    I'm sick of people requesting this absurdly expensive RAM. I've got not problem people requesting the RAM for over clocking, the good looking cooler mods, or the brand name... but people are still requesting this stuff because they expect much better performance.

    I can make my own Zero based graph based on your numbers to show people ... but I'm much rather direct them to the original source.
  • Wesley Fink - Wednesday, January 31, 2007 - link

    The explanation is bold and tells readers why we show a limited range. It is difficult enough to pick out memory variations in a reduced range chart, and it would be almost impossible in a zero-based chart. We also clearly state results on a zero-based chart would show little variation in performance at different speeds.

    In the conclusion I was also very clear:

    "There is also the reality that Core 2 Duo and AM2 really don't need the highest memory speeds to perform best. What they need is low latency DDR2-800 or possibly 4-4-3 DDR2-1067 to get the best performance possible. The unfortunate reality that we see again and again in memory tests is that the super high memory speeds are great for bragging rights and flexible overclocking, but they really don't do much for increasing actual real world performance on either the C2D or AM2 platforms.

    Memory companies seem obsessed right now with higher and higher DDR2 memory speeds... We wish memory companies would become just as obsessed with producing a moderately priced 2 GB DDR2-800 kit that can perform day in and day out at 3-3-3 timings. It would be a bonus if it also overclocked to DDR2-1067 with 4-4-4 timings. With Vista performing best with 2GB of memory many will be upgrading memory as they move to the new OS. Whoever finally produces a reasonably-priced low latency DDR2-800 2GB kit will sell all they can produce."

    We are reviewing super high-end memory that is a good choice for some buyers, like overclocking enthusiasts, but they are a waste of money for the average buyer. We definitely agree once you get above 800 on C2D or AM2 there is very little gain in performance. Your money is better spent on a faster processor or an upgraded video card.
  • Live - Thursday, February 1, 2007 - link

    If there is no real world difference then why do you state in your buying guides that it is important to keep a 1:1 ratio?

    From the January 2007 Buyer's Guide: Midrange Menagerie

    Using a 1:1 memory:bus ratio (which is typically optimal for overclocking)

    This is not the only example. You seem to do this in all your memory reviews I have read lately as well. Granted they have been of high end memory and it seems unreasonable to spend that much on memory if you weren’t looking for the absolute maximum speeds. I still don’t understand tough why a 1:1 ratio is good for overclocking. It would seems to me that by lowering your memory ratio you could reach higher CPU overclocks cheaper which is what really matters, no? I mean most overclock as a way to save money not get the WR in superPI.

    The thing is for us that don't work with this everyday it’s hard to keep up and not fall for the aggressive marketing. I would appreciate it if you could devote a little more time to steer us right without breaking the bank. I know it’s probably not as fun but it would be much appreciated by us readers.
  • Neosis - Sunday, February 4, 2007 - link

    I think additional latencies shall arise at chipset level. Generally following stages take place while requesting data from memory:
    1- Data Request from CPU thru. the FSB
    2- Chipset transfer the request to memory from FSB
    3- Memory responds to the request
    4- Chipset then trasfer the request from the memory thru. FSB
    5- Finally the requested data travels across the FSB to the CPU

    In stage two and four, running the memory asynchronous to the FSB will increase latency. Since conroe has short pipeline like Amd, memory latency is more important than the memory bandwidth.
  • tayhimself - Wednesday, January 31, 2007 - link

    Sorry but that is a piss poor excuse. Its like published scientific studies saying, "we use this statistical analysis because its the only one that shows a significant difference". Who the hell would buy drugs based on that premise. Piss poor AT!! Reply
  • thudo - Wednesday, January 31, 2007 - link

    Conroe E6700 @ 3.66Ghz (1.53V) || Scythe Mine CPU Cooler + ArcticSilver5 Thermal Paste || OCZ 2gb PC8000 DC EL GEEK RAM || Asus P5W DH Mobo (1301 Bios) || EVGA 8800GTX ACS3 Cooling Edition (659/1013) || 2x10k Rpm WD SATA2 Raptor HDDs || Thermaltake Armour Full Tower w/side 250mm Mobo Case Fan || OCZ GamerX 700W PSU || WinXP+ SP2 + DX9c + All Updates || Latest Nvidia Forceware (97.x) || 3dmark2006 Score: 12,426 || SiSandra Int Buff'd iSSE2 : 7561 MB/s // Float Buff'd iSSE2 : 7562 MB/s

    Still surprised my box beats out the latest ram with the Overclock. Can only imagine having the new ram overclocked on the beat! :O

Log in

Don't have an account? Sign up now