Gaming

Three DX9 games representing different gaming engines were used to test the performance of the Super Talent DDR3-1600C7 and the TEAM Xtreem DDR3-1600 in real world gaming. There are more recent gaming titles available, but they are also DX9. We will update games in the memory test suite as soon as a selection of DX10 games with reliable benchmarks is available. At that time the memory test OS will also be moved to Vista.

The Far Cry - River demo was run for 3 loops and results in FPS were averaged over the 3 runs.

Click to enlarge

Super Talent and TEAM dominate overall gaming performance of Far Cry and top the charts from 1333 to 2000+ memory speeds. Far Cry performs slightly better at 1066 with fast DDR2 memory on the P35 chipset, much as we expected.

Click to enlarge

Quake 4 and the underlying engine have always proved to be very sensitive to improvements in memory bandwidth. This is amply demonstrated in these memory tests. Again in all cases DDR2 and DDR3 are faster on P35 than the fastest DDR2 on P965. The pattern is the same as in Far Cry but the differences are magnified in Q4. The lower latencies of the new Micron Z9 DIMMs move that DDR3 to the top at 1333, and at every speed above it. It would be very interesting here if the CPU speed could be held constant and only the memory speed varied, but available ratios and the 333 multiplier make that all but impossible. What we need are 1600, 1666, and 2000 ratios available on upcoming motherboards - in addition to the 1333, 1066, 1000, and 800 that are already available.

Click to enlarge

We include Half-Life 2: Lost Coast as a representative of games that are less sensitive to improvements in memory bandwidth. Lost Coast is played through the Steam engine, where there is the constant worry, for a reviewer, that each new update of Steam will break your test benchmarks. Though the differences are very subtle and HL2 performance is most influenced by the video card used in the benchmark, there are nonetheless patterns that are exactly the same as the other two games. All P35 results are faster than the same fast DDR2 results on P965. DDR2 on the P35 is slightly faster at 1066 and the new Z9 DDR3 is tops in results at DDR3-1333 and higher.

The Super Talent and TEAM results are very close at all speeds, but Super Talent manages to be ever so slightly faster at every speed in every benchmark. Since the base memory chips are the same, a slight alteration in binning or small changes to programming the SPD could bring the Super Talent and TEAM into parity or move TEAM to the lead.

Bandwidth and Memory Scaling Conclusion
Comments Locked

31 Comments

View All Comments

  • TA152H - Saturday, July 21, 2007 - link

    There are a number of reviews that show power, but the numbers are pretty low in terms saving, because what you get from lower voltage you lose in higher speeds necessary to match DDR2 performance. I really should have said the extra performance doesn't come with the power penalty you'd expect, because the voltages are lower.
  • TA152H - Friday, July 20, 2007 - link

    I can see your point, because you are looking at it from a strictly system perspective, but his point is, it's a breakthrough from memory perspective. So, you're both right, you're just not talking about the same thing.

    Now, one thing that happens a lot with these reviews is their choices parts to test with the test item is questionable/poor, or incomplete. They end up with results that aren't as meaningful as they could be. I'll give a few examples, they benchmark mid-range and low-end DX10 cards, with Intel's high end CPU that would almost never be paired with it. Keeping in mind DX10 cards move work from the CPU to the GPU (or they should), it's not a meaningful benchmark and makes the cards look worse than they might with a lower end CPU (since this CPU is more likely to benefit from the offloaded work). Then they test the 1333 FSB processors with, of all things, 800 MHz DDR2 memory. Now, they test memory on a processor with 4 MB, instead of something smaller. This is not so bad, really, and I think he made a good choice because it is not too surprising to see this CPU, or any high end Intel CPU, with super fast expensive memory. But, if they tested it with a smaller cache CPU, you would probably see a much greater difference. So, I think both would have been helpful, because it's not easy to show a big memory difference when you're sitting on so much cache, although I think he made the right choice if he had only one test to make. But I think an additional one with a 2 MB or 1 MB processor might have given a better indication of what the memory can do, and I don't think it's meaningless because a lot of people will buy those and overclock them. Still, it would be more academic than real, but it would tend to add credibility to his point. This memory in some situations would have a really significant boost, but 4 MB of cache is hiding a lot of the performance improvement.
  • Wesley Fink - Friday, July 20, 2007 - link

    If you read the article you would see a long discussion of the difficulty of setting up the same CPU speed at different memory speeds. At present the only upper strap on P35 boards is the 1333 strap. We really need 1600, 1633, and perhaps even 2000 for more flexibility in USING and TESTING memory.

    Since you are so clever, why don't you figure out a set of multipliers and CPU speeds that can generate the SAME CPU speed at 1066 (or 1000) 1333, 1600 (or 1666), and 2000 using the 333 multiplier that is available in 1333, and having just a 1333 and 1066 strap. I can assure you we will use it if you can porvide a solution.

    I never mind criticism as it is how we imporve our performance, but I would appreciate it if you would actually read what is written before throwing rocks. That is common courtesy.
  • qpwoei - Saturday, July 21, 2007 - link

    Regardless of whether there is a solution or not, the fact remains that the numbers presented in this article are close to worthless with regard to comparing performance with respect to memory speed. The only useful numbers are the maximum speeds for the memory. Overall, this was a really poor article, and definately not up to the usual Anandtech standard (which is probably why people are laying into it so much).

    In particular the sentence:
    quote:

    However, few computer parts offer the kind of breakthrough performance advantage we see in these new DDR3-1600 kits.

    reeks of vendor press-release and almost made me write off Anandtech as being able to produce even halfway competent memory reviews. With Core 2, memory speed has next to no impact - you show it here in this very article, with the jump from DDR3-1066 to DDR3-1333 (a 25% increase in memory speed) only giving a 0.9% increase in performance. The *best* you're going to get from DDR3-2000 is a 3.2% performance increase, though in reality you'd be lucky to break 2.5%. "Few computer parts" - any CPU or GPU upgrade would get you a better increase in performance than the RAM.

    There's a couple other things which do not make sense as well in this regard. Why did you use 8x400 instead of 7x400? You could even run it at 7x400 and 6x400 to get an idea of the effect of CPU speed and get a rough idea of how it'd work at the "ideal" 6.7x400 setting.

    Ditto for 417 - Using 8x417 is pointless, testing at 6x would be closer and again testing at both 6x and 7x would provide an idea of where it fits in.


    So, how to fix it? My first suggestion to improving the tests is to bin the E6420, or at least only use it for corner cases. The 8x max multiplier is just too limiting. Either use a 266 MHz FSB CPU (E6600 for example), or use that X6800 that you've been using previously.

    Then figure out how the FSB:RAM ratios affect performance. I don't have this particular board, but from what I've read the available RAM:FSB ratios are 1:1, 5:6, 4:5, 3:4, 5:8, 3:5, amd 1:2, limited such that the RAM speed is not allowed to go below 200 MHz (eg: the 1:2 ratio is only allowed at FSB speeds of 400 MHz and higher). The configurations to test, then, would be:
    {} 200x10 = 2000 1:1 vs 250x8 = 2000 4:5
    {} 240x10 = 2400 5:6 vs 267x9 = 2403 3:4
    {} 267x6 = 1602 3:4 vs 200x8 = 1600 1:1
    {} 320x6 = 1920 5:8 vs 240x8 = 1920 5:6
    {} 333x6 = 1998 3:5 vs 200x10 = 2000 1:1
    {} 400x6 = 2400 1:2 vs 240x10 = 2400 5:6
    This will allow relative comparisons of all the ratios (the memory speed and CPU speed are identical for each pair compared).

    Assuming no significant differences show up above, you can move on to testing with non-1:1 ratios, which frees you up a lot. Even if there are slight differences, these can be somewhat taken into account later on. For example, your tests today could have been done with (give or take a MHz or two for rounding):
    {} 333x9 = 2997 4:5 (1066)
    {} 333x9 = 2997 1:1 (1332)
    {} 500x6 = 3000 4:5 (1600)
    {} 500x6 = 3000 5:6 (1667)
    {} 500x6 = 3000 1:1 (2000)

    Even with your less-than-ideal E6420, you could have done:
    {} 444x8 = 3552 3:5 (1066)
    {} 444x8 = 3552 3:4 (1332)
    {} 500x7 = 3500 4:5 (1600)
    {} 500x7 = 3500 5:6 (1667)
    {} 500x7 = 3500 1:1 (2000)

    This is just from messing around in Excel for half an hour. With a bit more thinking, there's probably some even better options.
  • Wesley Fink - Saturday, July 21, 2007 - link

    Thank you for your suggestions. Your scheme is better than the variation from 2.66 to 3.5, but it still will not be apples to apples. We have already compared high multipliers like 500 to equivalent speed at say 333 and there is a significant difference in our test results - minor in some tests and major in others. Our amplification article looking at the FSB speed increase component and the P35 component in the P35 performance improvements examined this. That was comparing 266 to 333 - the difference in 333 to 500 is even greater. Ideally all memory test speeds would be at the same FSB and multiplier.

    I ran your numbers through the BIOS and at 333x9 the ratios provide 800, 1067, and 1333 memory speeds. At 500x6 you can select 1000, 1200, 1600, and 2000. No 1667 available, but the 1067 and 1333 would be at much slower FSB speeds even though the 333x9 CPU speed is the same as 500x6. 333 is really 333.3333... We are ignoring the fact that 1:1 is the preferred ratio, but as we have shown in the past, that is not really a big concern as the impact is really minimal on the Intel chipsets - much smaller than you would expect.
  • qpwoei - Saturday, July 21, 2007 - link

    What BIOS version are you running on that board? The 2.58 BIOS as shown at
    http://www.octech.jp/modules/wordpress/index.php?p...">http://www.octech.jp/modules/wordpress/index.php?p...
    shows a much better range of ratios.

    I'd forgotten about the "Intel P35 Memory Performance: A Closer Look" article, so I went back and had a closer look. I was really surprised about the results in there - and then I remembered that the Core 2 FSB bandwidth is only half the memory bandwidth and it made sense again :) I think that trying to benchmark memory on a platform with a bottleneck between the memory and CPU is going to be very difficult.

    It looks like you're basically hosed, then, apart from the ratios that Intel provide. Increasing the CPU speed warps the results, increasing the FSB speed warps the results, so all you've got left is the FSB:RAM ratio, which really doesn't help you too much. Personally, I wouldn't bother with the "real world" results (ie: simply stick to the synthetic memory benchmarks) since they don't really add anything to the article. However, it would be interesting to see one article with a benchmark at each of the memory ratios offered by the P5K3 (keeping FSB and multiplier constant).

    I've got a few more ideas, but I'll just email you directly to save cluttering up the comments too much. Also, I'm probably going to be picking up a P5K3 (and a Q6600 like world + dog after the price cuts) so I'll probably be doing a bit of investigation on my own :)



    Oh, for the easy days of memory benchmarking on the K8 ...
  • Wesley Fink - Monday, July 23, 2007 - link

    Let me clarify a bit. THese ratios are available when the FSB strap is set to Auto. That means at the various memory speeds the strap may be 200, 266, or 333. We want the strap constant at 333 if at all possible for consistent results and comparisons.

    The issue of the memory strap probably deserves a separate article for those not familiar with this concept and its impact on memory performance.
  • Wesley Fink - Monday, July 23, 2007 - link

    The additional ratios shown on the Japanese website are only available if an "Auto" setting is used for FSB.
  • LTG - Friday, July 20, 2007 - link

    I didn't say the test setup was easy, maybe it's not even practical right now to realistically measure the gains of this memory.

    However until it is able to be measured well, do you think it's appropriate to conclude that this memory "offers breakthrough performance advantage" ?

    You ask for constructive criticism. I would suggest that you not issue such grand conclusions until it's possible to perform benchmarks that support them.

    Is that throwing rocks?

  • retrospooty - Sunday, July 22, 2007 - link

    "I would suggest that you not issue such grand conclusions until it's possible to perform benchmarks that support them. "

    Perhaps you are overemphasizing the word "breakthrough". It CAN mean a revolutionary new way of doing things, or discovery, but it can also mean a barrier having been surpassed... for example 1600mhz was not possible with DDR2, thus its a breakthrough of sorts. Another example is the latest 1TB hard drives. Hardly revulutionary, but you could call them breakthrough for surpassing the 1TB mark.

Log in

Don't have an account? Sign up now