morphine wrote:
I've seen plenty of people refer to that article as evidence that there's no point in going higher than 1333 memory for Sandy Bridge. I disagree.
I'm not saying that it's necessarily the case that faster memory will give a significant benefit, but simply that the article doesn't show that the opposite is true in the general case.
The article simply shows that under the conditions used, there is no significant benefit for the majority of the performed tests. This is not the same thing.
First of all, not all test showed insignificant gains from increased memory speeds. The Euler3d test actually saw quite a big improvement, which means that there *are* applications where the increase in memory speed can be useful. However, this actually wasn't the test that saw the biggest improvement. The biggest improvement (by far) when going from 1333C9 to 1600C9 was actually seen with the minimum frame rates for Metro 2033 (High). I'm not sure how much this test can be trusted though, since the 2133 memory showed slower results than the 1600 memory. Yes, the memory timings were looser for the 2133 memory, but the increased frequency should offset this. This, and the somewhat weird minimum values for Dirt 2 (High) and Metro 2033 (Low) make me wonder how reliable these minimum frame rate tests really are (3 out of 4 tests seem potentially "weird").
It should also be noted that many of the other tests don't really have a high enough precision to reliably show a difference on the order of 5% or so. For example, the Civ 5 (High) test shows a 5% improvement when going from 1333C9 to 1600C9, but this is only a difference of 2 fps, which makes it harder to draw any reliable conclusions (rounding alone could have a big impact), and this could also be hiding improvements in other cases.
In addition to this, there are no tests of heavy multitasking conditions, a case where I'd expect increased memory speeds to be helpful.
Aside from the questions of which tests are (not) present in the article, and how representative these tests are for what you're likely to use your computer for, there's also the question of what happens when you work outside the parameters of the article. For instance, as Airmantharp says, there's the case of overclocking. The Sandy Bridge processors are good overclockers. In addition to this, there's also the possibility of using a faster graphics card. The test uses a 5870, and there are certainly faster cards out there even now (Geoff even states in the article that the graphics card becomes a bottleneck in some cases, and when the graphics card is the major bottleneck, you can't really draw too many conclusions about memory speed (other than "in this particular scenario, it doesn't really help")). It's also possible that you'll want to put a next generation card into your computer eventually (I certainly am planning to do that). When you increase the speed of "everything else", memory becomes more of a bottleneck.
It is of course possible that 1333 memory is fast enough to not become a major bottleneck in many of the tested situations even if you overclock your processor and use a faster graphics card. Unfortunately, the article doesn't show if this is the case or not. It would have been interesting to know how far we are from memory being a bottleneck in the tested scenarios. Would 1200 MHz memory be fast enough to avoid being a bottleneck? 1000 MHz? 800 MHz? If memory wouldn't be a bottleneck (for a given scenario) at 800 MHz, it would probably be safe to say that 1333 MHz memory wouldn't easily become a bottleneck when you overclock your processor either. At the very least, it'd be very interesting to see what the situation is like with an overclocked processor and a faster graphics card.
We also have a wild-card in AVX.
By taking advantage of AVX, it should be possible to increase the demands on the memory sub-system. It seems likely that during the lifetime of a new Sandy Bridge system, *someone* should be able to release *something* that would be able to make good use of AVX. Would faster memory be more useful in a larger number of cases then? Who knows...