Memory Performance

We utilized Memtest86+ and our full benchmark test suite to optimize our DDR2-667 and DDR2-800 memory for both performance and stability. Our overclock settings represent how far we could overclock our memory modules with the best possible timings on the Gigabyte DS3 motherboard. These settings had to pass our full benchmark test suite and not just the applications that we are showing today. The results on other chipsets such as the Intel 975X or NVIDIA nForce 590SLI can and will be different due to chipset timings and memory SPD compatibility. Our test results utilizing the GEIL DDR2-800 memory from our P965 roundup are listed at the bottom of the chart for comparison.

Click to enlarge

The results are somewhat surprising with the A-DATA DDR2-667 memory reaching DDR2-970 at 5-5-5-18 timings. This represents a 45% increase in memory speed that only required an increase in voltage to 2.2V. We were able to POST and enter Windows XP at DDR2-1000 with this memory but it was not stable enough to complete our benchmarks. Our PQI Turbo DDR2-667 memory was able to reach DDR-940 with 2.2V resulting in a fairly impressive 41% overclock at 5-5-5-15 timings.

The Transcend JetRam DDR2-667 and G.Skill DDR2-667 were both able to reach DDR2-910. This represents a 36% overclock and is what we expected out of our memory choices on average. The Transcend JetRam memory was listed at $76 per 1GB module at ZipZoomFly when we purchased ours, reminding us of the "good old days". We did a last minute check today and it is sold out now. However, if you can find this memory in stock then buy it as it offers the best price to performance ratio of our test candidates today.

There are never any guarantees in overclocking and our Corsair Value Ram represents this premise. We could not get the memory to run at DDR2-800 at the 4:5 ratio and could only reach DDR2-750 in our overclocking results. This required a change to 5-6-5-18 timings that resulted in performance just slightly better than our stock DDR2-800 settings. We have seen the same results on our AM2 systems so beware that this memory runs at its advertised settings and not much more.

Our GEIL DDR2-800 memory was only able to reach DDR2-910 (7x455FSB) due to the Micron D9 1GB issue that has been plaguing the Gigabyte GA-965P-DS3 motherboard. We are glad to report that Gigabyte has identified this issue and hopefully will have a new BIOS available in the near future. Our GEIL DDR2-800 memory reached DDR2-1040 (7x520FSB) on our Asus P5B-E 1.02G motherboard so we know it has the capability to overclock further than our DDR2-667 selections. However, this only represents a 30% overclock of the memory compared to our ability to overclock most of the DDR2-667 selections from 36% to 45%.

Overall, our DDR2-667 memory selections, sans the Corsair Value Ram, performed admirably and scored within a single percent of our much more expensive DDR2-800 memory at stock settings. On our P965 motherboards we generally see that system performance at DDR2-667 memory speeds with timings around 4-4-3-10 or better is perfectly acceptable. Due to improved overall system performance we still recommend DDR2-800 memory speeds with timings at 4-4-4-12 or better when not overclocking the system. The difference in timings between 4-4-3-12 and 3-4-3-9 at DDR2-800 is negligible and in our opinion is not worth the extra cost. We would have to recommend the purchase of the Transcend memory based upon our results as their price to performance ratio cannot be beat in this category. We will present budget DDR2-533 and DDR2-800 memory results in the near future.



View All Comments

  • Gary Key - Friday, October 27, 2006 - link

    I can already tell you that our mATX board roundup will include a lot more AM2 boards than Intel at this time. We finally have some good BIOS releases on a couple of our Core 2 Duo capable mATX boards that are worth testing now. Still do not know when NV will release an updated NF61S board for C2D. Reply
  • kmmatney - Friday, October 27, 2006 - link

    So, what gives the best bang for buck for overclocking?

    Option 1) Slower E6300 cpu with more expensive memory that overclocks better

    Option 2) A more expensive E6400 with a higher multipler, and slower, cheaper memory?

    It sounds like your better off getting a more expensive cpu with a higher multiplier, instead of more expensive RAM.
  • JarredWalton - Friday, October 27, 2006 - link

    I would typically go with the E6400 over the E6300 as it offers more flexibility overall. Hmmm... maybe that's why I bought one? :) 8X vs. 7X helps a lot. 3.6 GHz requires DDR2-900 and 450 FSB with the E6400, which is pretty reasonable. The same CPU speed with E6300 would require RAM that can run at over DDR2-1000 (DDR2-1028 to be exact). *MUCH* more expensive for that sort of RAM. Reply
  • xsilver - Friday, October 27, 2006 - link

    what about the option of using ratio's on a e6300 and using the cheaper ram?
    what kind of performance hit is to be expected?
  • kmmatney - Friday, October 27, 2006 - link

    Thanks. Makes sense to me. Reply
  • yyrkoon - Friday, October 27, 2006 - link

    Heya Jarred, according to a buddie of mine, even multipliers on CPUs is prefferable. He claims that odd CPU multipliers is in efficient, and actualy underperforms even numbered multipliers . . . I've NEVER heard of this before, but this doesnt mean its not true, and I've never done any testing here myself.

    Also, I was curious , when does memory timings make a real world difference ? Memory SPEED (ie running 1:1 vs any other ratio ) I would think would be the biggy. Here, I've done some testing, and to be honest, outside of 'benchmarking for a living' I honestly havent noticed much, if any difference.
  • Aikouka - Friday, October 27, 2006 - link

    Timings tend to make a difference in almost any memory hungry application. However, the one thing to keep in mind is that not all memory timings will have a great effect or any effect at all. The two main values that are best taken into consideration are CAS latency and Command Time (not usually shown when buying memory). CAS latency is always taken into consideration, regardless of what memory address you're accessing. Command Time can generally speed up your system as it defines how long a command should be sent to the memory module for. If it's lower, then another command can be sent sooner.

    Most of the other timings deal with rows and how long it takes to switch from getting a row into the buffer to finding the column (tRCD), how long for a row to be brought into the buffer (tRP) and how short can the duration be of a row sitting in memory (tRAS).
  • JarredWalton - Friday, October 27, 2006 - link

    Regarding even vs. odd multipliers, it's not more or less efficient, but merely an impact on memory speed... but not with Intel or Core 2 Duo. AMD's socket AM2 (and 939 and 754) derive memory speed from the CPU speed, so a 3000+ AM2/939 chip has a 9X multiplier and a 200 MHz base HyperTransport bus speed, giving 1800 MHz. Back in the DDR world, you also had PC-3200 RAM running at a 200 MHz bus, so multipliers didn't matter. With AM2, things change a bit.

    The same 3000+ has a 9X multiplier, but where a CPU/9 divider would give DDR2-400, that's now slow memory. If you want DDR2-533, how do you divide 1800 by a whole number to get a 266 MHz base bus speed? Answer: you can't. So you come close: 1800/7 = 257, so if you have a 3000+ and you set the RAM for DDR2-533, you really get DDR2-514. It's doubtful that you would notice the loss of 19MHz of RAM bandwidth, but it's still there. For DDR2-667, you also can't evenly divide 1800 to get 333, so you end up with CPU/6 = 300, or DDR2-600. In that case you lose 67MHz of RAM bandwidth. Finally, DDR2-800 ends up using CPU/5 = 360 = DDR2-720, a loss of 80MHz of potential bandwidth.

    Now, here's why even multipliers are "better". You can always get DDR2-800 memory speed with an even multiplier, as you simply divide the CPU speed by half the CPU multiplier. 2000 MHz (X2 3800+ or 64 3200+) uses a 10X multiplier, so CPU/5 = 400. 2000 also happens to work well with DDR2-667 as CPU/6 = 333, so you end up with exactly DDR2-667. For DDR2-533 you get CPU/8 = 250 or DDR2-500.

    Not all even multipliers work exactly with DDR2-667, but they all will give you DDR2-800. The 2400MHz chips also give DDR2-533 exactly via CPU/9. In general, I think even multipliers have a better chance of coming close to the rated memory speeds, but it's really not a huge issue as AMD isn't typically bandwidth constrained.

    As for Intel, they derive memory speed from the bus speed using ratios rather than going from CPU speed, so with a 266MHz base FSB you always have 1:1 giving DDR2-533, 4:5 gives DDR2-667, and 2:3 gives DDR2-800. You can also go 1:2 for DDR2-1066 if the motherboard supports it.

    If the memory timings are equal, in almost all cases higher bandwidth will be equal or better. Core 2 does make one small change, however, as if you overclock past a 500 MHz base bus speed (I think that's right) the chipset bootstrap changes. Normally, the chipset runs at the FSB speed, so overclocking improves chipset performance as well. At 500MHz the chipset reduces speed assuming you're using 1333FSB instead of 1066FSB. Something like that - Gary could explain it better. Basically, there's a point at which the chipset speed is reduced and so even if the bus is 20 MHz faster you might actually get slower performance. Without extreme overclocking, this is rarely a concern.
  • imaheadcase - Friday, October 27, 2006 - link

    From looks of prices of might of found something more expensive than ram for a "budget" system :P Reply
  • Madellga - Friday, October 27, 2006 - link

    Hi Gary, nice article as usual.

    The irony is that prices started to drop in Europe. They surged 50% and more between August and end of September, but there was a substantial price drop in the last 10 days: from 15 to 30%, depending on the brand and speed.

    Take care.

Log in

Don't have an account? Sign up now