How We Tested

If there was one thing we strongly suspected at the outset and know for certain now, it’s that no single memory benchmark has the breadth or depth to give us an accurate representation of any particular module’s attributes. Much like the programs we run on our PCs, some benchmarks favour bandwidth while others favour latencies. A memory module that blitzes its way to the top of the pack in, for example, PCMark04, may actually trail by a couple of frames per second in your favourite game title.

So how did we try to counter this? Well, we did it by collecting eight individual results from a suite of seven different benchmarks, each run a minimum of three times, and then we averaged the results to give an overall rank.

Now before all the boffins among you start formulating your derisory emails, don’t! We fully appreciate that this is just one of probably many hundreds of ways that the testing could have been carried out, including the employment of far more technical software or hardware allowing more advanced data error testing, but this isn’t a mission critical server memory roundup, it’s a memory roundup for the guy or girl on the street who wants to know what’s best and at what price.

Our primary interest in this particular roundup was how the various memory modules performed when run in dual channel mode, and the reasoning for this is simple; the vast majority of the motherboards that will be sold this year will be based on dual channel memory controllers. This is also why we decided not, on this occasion, to see how the different makes performed when run as single sticks. While this would almost certainly favour memory not optimised for dual channel operation, it wouldn’t necessarily prove as beneficial for memory that is.

Testing was carried out on two platforms, or more accurately on one platform with two different CPUs fitted. All the results that follow were generated on an Abit AI7 motherboard run off a Tagan 480W power supply with either a Prescott 3.2 GHz CPU or a “Northwood” 2.4C CPU installed - the former to show how Prescott has improved on overall memory performance and the latter because we know it overclocks to at least 280MHz FSB.

Interestingly, we assumed that most if not all of Prescott’s improved memory performance was due to its increased cache levels, but the un-cached read results suggest that Prescott is simply a more efficient architecture when it comes to memory transactions.

In an attempt to push a little further we also tried running an unlocked 2.4C alongside a “modified” AI7 BIOS which allowed for multiplier adjustments but unfortunately it wasn’t able to keep pace with our locked processor.

Torture testing was carried out on each module with a two-hour memory burn-in performed both at 200MHz and at the maximum stable overclocked frequency just to confirm that operation at these speeds was likely to be trouble free. If a module failed it was first re-run at the same frequency before having the frequency lowered if it failed a second time. Sufficient time was given for module temperatures to return to normal before each re-run. Those of you who hold with the theory that memory needs time to “burn-in” before it overclocks properly, whether it’s a myth or otherwise, will be pleased to know that the torture testing was done before the Overclocking, despite the risks of completely destroying them in the process and the possibility we’d have nothing to test at the end of it all.

Overclocking was carried out under a low ambient room temperature of 18c and then again under a raised ambient temperature of 28c just to try and identify whether or not heat plays a factor in overall reliability. This may seem a little unfair but even at 28c we’re probably still below the kind of temperatures encountered inside most closed-case PCs.

Overclocking was also carried out at 2.6, 2.7, 2.8 and 2.9v in all cases other than where the module’s rated voltage was higher than this, in which case we continued until that voltage was reached.

So that’s the basics covered, and before you go asking why we didn’t run tests at every one of the dozens of possible timing combinations available, or test them on every make of motherboard known to man to verify which ones accurately interpret SPD settings, or paint them white and see if this helps reflect solar radiation spikes caused by sunspot activity, it’s because it would simply be too time-consuming, not to mention expensive. Having said that, if enough people want a particular test at particular settings, a bit of a follow-up isn’t impossible.

And finally on the subject of timings, we decided for our sins to work to the lowest common denominator and run all the overclocking tests at the same timings as those used by the most lax modules on test to try to even the playing field a little. This of course meant testing all CL2.0 modules at 2-4-4-8, CL2.5 modules at 2.5-4-4-8 and CL3.0 modules at 3-4-4-8. Not a perfect scenario we agree but perfectly valid for our tests, even though you may well be able to squeeze out better numbers by tightening the settings a little

comments powered by Disqus