Influence on performance, memory size

Influence on performance

So what is the influence of a fast or less fast DDR2 memory on performance? This is something we wanted to look into with AMD as well as Intel, using two programs that are sensitive to memory bandwidth ; Far Cry and WinRAR, the file compression software.

To start with, here are some results with an Athlon 64 X2 at 2.6 GHz equipped with 1 MB of cache for each of its cores. It should be noted that at this frequency and because of AMD’s architecture of its memory controller, DDR2-800 functions at only 93% of this frequency versus 98% for DDR2-667 and DDR2-533.


As you can see, the AM2 platform is very sensitive to its accompanying DDR2 memory, and more so to frequency than timings. Thus, a DDR2-800 memory with large timing values is a better option than a DDR2-667 with more aggressive timings!

Here are the same figures, this time with a Core2 Duo E6600 at 2.4 GHz.


The impact is less significant if you compare with memories with lower performances. Contrary to the AM2 platform, timings have more of an impact than frequency, but this doesn’t stop DDR2-1067 in 5-5-5-15 from being faster than DDR2-800 in 4-4-4-12. This is true even if we have to admit that part of the gain is related to the fact that the i975x chipset has optimal performances when the ratio of memory and bus processor frequencies are whole numbers, here at a value of 2.

On the Intel platform, the bus system frequency is moreover a rather important point, because the most widespread chipset for the Core 2 Duo, the P965 Express, doesn’t allow memory to function at a lower frequency than that of the bus processor’s. With current processors, which function in FSB1066 (a real 266 MHz), we should at least have DDR2-533. With future ones in FSB1333, you will therefore need DDR2-667, and this could easily be more.

Indeed, it’s common to use FSBs of 400 MHz or more for overclocking, and at this frequency DDR2-800 memory is an absolute must as well as good timings. Even on a chipset that offers « lower » desychronization, it will always be of more interesting to have DDR2-800, because DDR2-800 in 5-5-5 will be as good as DDR2-600 in 3-3-3.

So, even if in terms of performances it’s only on the AM2 platform that DDR2-800 is really of extra benefit outside of overclocking, the Core 2 Duo platform is capable of using this memory and even requires it in overclocking. Of course, you shouldn’t go in the opposite direction and buy certain modules with very high certifications and amazing numbers while their interest in practice is very limited.

And the quantity of memory ?

PCs need more and more memory. While a short time ago we were Ok with 1 GB for intensive use, 2 GB are now strongly recommended, and this even with Windows XP ! This doesn’t even take into account large video editing projects or 3D rendering in heavy scenes in games like Battlefield 2142 or Supreme Commander, which need 1 GB alone or even 1.5 for the latter ! To not have 2 GB equals long loading times and also unpleasant cuts in fluidity.

The arrival of Vista also means additional needs in terms of RAM, first of all, just for its new functions such as the Aeoro interface or the indexing of files. In addition, SuperFetch technology, which pre-loads the most commonly used data and programs into memory to have added reactivity, will not be too effective if enough memory is not available. In short, the Vista experience will only be better with 2 GB of memory.

Of course, for less intensive uses 1 GB can be sufficient with Windows XP and even Vista, but entry level 512 MB configurations should be avoided like the plague. On the other end of the spectrum, having 4 GB of memory is fortunately only for specific uses, and all the more so because the 32 bit Windows version is limited to this size and a maximum of 2 GB per application. If the application does support it, you can raise the limit to 3 GB by modifying the initialization file, after this you will have to make the transition to 64 bits.

The comparison

The comparison

With the plethora of models offered by manufacturers, we limited ourselves to a specific area, kits with two chips of 1 GB PC2-6400 (DDR2-800) modules certified at timings of 4-4-4. This gives us a total of ten products, with one exception, the G.Skill module certified at 4-4-3, which is a slight difference of small influence.

Here is the exact list of modules tested with a sample of prices from 04/18 from the three retailers that supplied us with the kits and rounded to the nearest 5 euros. It should be noted that recently the price of DDR2 has fallen quite a bit and should do so even more in the following weeks.

- A-Data ADQVD1A16 (250 € at PC Look)
- Corsair TWIN2X2048-6400C4 (220 € at LDLC, 220 € at Materiel.net, 240 € at PC Look)
- Corsair TWIN2X2048-6400C4D (245 € at LDLC, 245 € at Materiel.net, 250 € at PC Look)
- Crucial BL2KIT12864AL804 (245 € at LDLC)
- G.Skill F2-6400CL4D-2GBHK (200 € atMateriel.net)
- G.Skill F2-6400CL4D-2GBPK (185 € at Materiel.net)
- G.Skill F2-6400PHU2-2GBHZ (220 € at Materiel.net)
- GeIL GX22GB6400UDC (180 € at PC Look)
- Kingston KHX6400D2LLK2/2G (240 € at Materiel.net, 220 € at PC Look)
- OCZ OCZ2P800R22GK (190 € at Materiel.net, 190 € at PC Look)

This wasn’t a question of just pure performances, because quite simply under the same parameters all modules offered strictly the same results. For each chip, we therefore tried to find the maximum settings with three rather standard latencies : 3-3-3-8, 4-4-4-12 and 5-5-5-15. Tests in 3-3-3-8 were done with a FSB:DDR ratio of 1:1, versus 1:2 for the others, in order to not have any problems related to increases in FSB.


To validate stability, we used Memtest86+ 1.7 in its standard configuration. For frequency combinations, timings and power use, results were validated when the chips didn’t have any errors in two passes with our test motherboard, an ASUSTeK P5B Deluxe. Given that each trial lasted roughly 30 minutes and that there were 10 kits tested in over 160 settings, we hope you understand why we didn’t choose 10 passes.


All chips used in this test come from store shelves in order to avoid specially prepared modules for the press from dishonest manufacturers.


Of course, there is always a bit of randomness in results with overclocking. All the same, we noticed during tests that several models of the same kits only differed by a max of 10 MHz. The chips didn’t have direct ventilation, but rather a classic airflow characteristic of PCs that are well ventilated.

A few particularities

All chips look more or less the same, however, some manufacturers tried to make theirs stand out. Corsair, for example, offers models with classic radiators on top, diodes that light up, or its Dominator version with an imposing cooling system devoted to cooling down the module if a light air flow is present. Crucial’s Ballistix Tracer is equipped with activity diodes as well as LEDs at the base of the chips that light up the motherboard’s PCB. This is something for tuning fans!


For official specifications, they are all quite close with two exceptions. The G.Skill HK offers timings of 4-4-3-5 and the OCZ is at 4-4-4-15, versus the standard 4-4-4-12. OCZ’s difference doesn’t have a real measurable impact on performances, and in practice the frequencies in 4-4-4-15 and 4-4-4-12 are the same. For the 4-4-3-5, results are only slightly different compared to the 4-4-4-12 (0.5%). We would have preferred chips certified in 3-3-3, but unfortunately the kit that comes the closest to this, the Corsair Dominator TWIN2X2048-6400C3DF, costs more than 500 €. We’ll have to pass.


Besides timings, there is of course the PCB as well as the chips themselves that determine the quality of the module. However, knowing that these products come directly from stores, we decided not to take off the radiators to identify the chips. Let’s allow test results tell us which ones can go beyond official specifications.

2 kits had problems

Even if they are just details we wanted to tell you about two small problems we encountered. The first OCZ kit we received was defective and showed errors at every frequency. Everyone knows this can happen, but buying a kit that is certified at certain frequencies should avoid this type of inconvenience.


The second problem came from G.Skill’s HZ kit. As we finished all tests for each module in 4-4-4-12 at 2.45V to see if this added anything extra, it showed an error message and also the chip became irreparably damaged. We obtained two new kits, which didn’t have this problem and they showed the exact same performances as the first.

Results in 4-4-4-12

Results in 4-4-4-12

We start with results in standard timings. As you can see six models reached 460 MHz at only 2.1V, and five of them attained 490 MHz at 2.25V (the A-Data, the 2 Corsairs, Crucial and the G.Skill HZ).

Three kits go as far as 510 MHz but with voltage from 2.35V to 2.45V. Just to remind you, DDR2 official specifications are at 1.8V, and for the long term we don’t recommend you going beyond 2.25V or even 2.35V with good ventilation.

At the back of the pack, there are the kits limited to 440-460 MHz whatever the voltage: the G.Skill PK, GeIL, Kingston and OCZ. If the PK is the only one stuck at 440 MHz, it is however at a normal voltage of 2V. Of course, we are beyond official specifications even at these frequencies, however, given the performances of other kits we could have expected a little more.

3-3-3-8 and 5-5-5-15

Results in 3-3-3-8

Here are stable frequencies obtained with very aggressive latencies of 3-3-3-8. The differences are even more pronounced than in 4-4-4-12, because if four kits reached 370 MHz in 2.25V (the A-Data, Corsair Dominator, Crucial and G.Skill HZ), three others don’t go over 310 MHz at this voltage: the GeIL, Kingston and OCZ, which were already last in 4-4-4-12.

The G.Skill HK reaches the highest, but the voltage necessary is at 2.35V from 370 MHz and 2.45 at 400 MHz. This is good for sporadic use but not in the long term.

Results in 5-5-5-15

The change to 5-5-5-15 gives noticeably higher frequencies for all models. It is, however, the Crucial and G.Skill HZ, which both stand out with 580 MHz, however at 2.45V for the first. This is no less than 60 MHz more than the kit in third place, the Corsair Dominator.

In the 490-510 MHz slot we find the seven others and once again the GeIL, Kingston and OCZ and in the back of the pack. The first two immediately need a voltage of 2.25V despite modest frequencies, and the last is stuck at 490 MHz whatever the tension.

Conclusion

Conclusion

We will start with the worst, which is something we never do. The GeIL, Kingston and OCZ respect initial specifications, don’t go too much higher, and are often last in tests, especially with aggressive timings. While the GeIL and OCZ modules make up for this because of their low price, the Kingston doesn’t even offer this, and is therefore of little interest.


In terms of performance, two kits really stand out; the Crucial and G.Skill HZ. Whatever the timings, they always finished in the lead. At 220 €, the G.Skill HZ also have the advantage of being in the average price range for this type of module. The Crucial is a little more expensive, but the counterpart is that they have a more attractive design with the activity diodes and the ones that light up the PCB of the motherboard. This is of course a negligible point for some.


Amongst those priced at around 200 €, once again the G.Skill comes out on top with its HK model. It reaches very good frequencies (except with an average result in 5-5-5-15) and needs slightly more voltage than our two preferred models. These good results overshadow the more expensive modules with equivalent performances, the A-Data, Corsair or Kingston. As a reminder, following the large drop in DDR2 prices in recent weeks, this should be even more pronounced with decreases of 10 to 20% in the next few weeks.

Overall, if there are three kits that are preferable, none are to be avoided like the plague. Of course, you can always have isolated problems like we had, but generally specifications are respected and they are surpassed. We finish this article with a warning regarding the validity of these test results in the long term. Some manufacturers interchange chips or even brand names incognito, while official specifications remain the same. Luckily, the Crucial and G.Skill HZ, which finish first, are known to stay with Micron chips (to which Crucial is a subsidiary) of the D9 line, which are currently some of the best.

Update : The announced price drop has already started. Before the cheapest model was the GeiL at 190 €, and now it is the OCZ and G.Skill HK starting from 160 €. You may recall that this last kit was in our top three at 200 €.

Our two co-winners, the G.Skill HZ and Crucial Ballistix Tracer will also go down, because they were from 220 to 245 € and now are at 195 to 170 € ! At this price the Crucial is by far currently the best choice.

These differences are of course related to various delays in price decreases. In the next few weeks most should be found at 150-170 €.