The cards, DDR3

The motherboards

For this test, we used three motherboards built around this new P35 Express chipset. The first two are designed by ASUSTeK, the P5K Deluxe and P5K3 Deluxe, which manage DDR2 and DDR3, respectively. The Gigabyte P35C-DS3R motherboard, on the other hand, allows the use of two DDR3 slots as well as four DDR2 slots for a more gradual transition.


In terms of functions, ASUSTeK has added another Gigabit port managed by a Realtek PCI chip to those already available in the P965 Express. We should note, however, that the first port isn't managed by the chipset, but rather by a Marvell PCI Express chip. FireWire is handled by an Agere controller, while an Analog Device AD1988B serves as the HD Audio codec. Finally, ASUS made the nice move of integrating an IDE/SATA JMB363 PCI Express controller chip, which can manage one classic IDE port as well as two e-SATA ports, a WiFi Realtek chip, and USB.


The Gigabyte card, which is not in the same class, has fewer additions. The Gigabit network of the P35 doesn’t seem to excite much emotion because Gigabyte uses a Realtek PCI-Express chip, the same manufacturer supplying the HD Audio codec, an ALC889A. The Jmicron controller is also part of the action, here for the internal SATA which reaches the 8 ports.


Besides their chipset, these motherboards share the particularity of having the latest power supply specifications for processors, giving them assured compatibility with the upcoming Core 2 engraved in 45nm and known as the Penryn. All P35 Express motherboards should theoretically be compatible. For previous motherboards, this may not be the case, but the limitation isn’t related to the chipset, and it is possible to have motherboards based on older chipsets come with « Penryn » revisions.


In terms of performances, given the current state of things the Gigabyte is equivalent to the ASUSTeK motherboards with DDR2. However, in DDR3-1333 mode, we noted that it was behind compared to the P5K3, while being equal in DDR3-1067. In addition, this motherboard showed it was not as good in overclocking, not being able to reach a 450 MHz of FSB with stability. Actually, we only used it to measure the impact of DDR3 on power use and all other tests for performances were carried out on the P5K and P5K3.

DDR3

Compared to DDR2, the main changes related to DDR3 are power (which goes to 1.5V versus 1.8V in DDR2 and 2.5V in DDR) and prefetch. The latter, which went from 2n to 4n bits in the transition from DDR to DDR-2, now increases to 8n bits. The internal organization of memory cells was therefore modified to have a doubled transfer rate without raising the frequency at the expense of that of entry/exit buffer and the external memory bus.


Besides the lower power use of DDR3, this memory should bring a noticeable gain in terms of bandwidth, because it will be officially available in the long run from DDR3-800 to DDR3-1600 versions. This is the double of DDR2 which goes from DDR2-400 to DDR2-800. Of course, like for DDR and DDR2 certain chips should go further than official specifications, and we will surely see chips functioning at DDR3-2133, which corresponds to PC3-17000 chips.

In terms of latency, however, the DDR3 doesn’t add anything, because with Micron, for example, according to chips we will be at the following intervals :

- DDR3-800 : 5-5-5 to 6-6-6
- DDR3-1066 : 6-6-6 to 8-8-8
- DDR3-1333 : 8-8-8 to 10-10-10
- DDR3-1600 : 9-9-9 to 11-11-11

These are of course clock cycles and all of this is also dependent upon frequency. A DDR3-1600 in 10-10-10 won’t have slower access than a DDR3-800 in 5-5-5, because it is clocked twice as fast. It should be noted that a DDR3-800 chip in 5-5-5 should have access that is slightly slower than a DDR2-800 in 5-5-5, as memory cells function internally at 100 MHz for the first and at 200 MHz for the second.


For this test, Corsair was able to provide us with its first DDR3 modules, the CM3X1024-133C9DHX. Still in their sample state, the chips we had in our possession are specified to function in DDR3-1333 with latency of 9-9-9-24.

In practice, here are the stable timings which we were able to obtain on these chips with the P5K3 :

- DDR3-800 in 5-5-5-15 at 1,8V
- DDR3-800 in 6-6-6-18 at 1.5V
- DDR3-1067 in 7-7-7-20 at 1.5V
- DDR3-1067 in 8-8-8-22 at 1.5V
- DDR3-1333 in 8-8-8-22 at 1.9V
- DDR3-1333 In 9-9-9-25 at 1.5V

As you can see, we were able to go down a level in comparison to official specifications with variable voltage. We noticed, however, that Qimonda's chip didn’t seem to give us the highest performance in DDR3, because, for example, other chips already allowed a functioning in DDR-3 1067 in 6-6-6, which was impossible with these chips. We can be sure that Corsair will use them on its Dominator, the DHX being the « classic » model.

Update

Kingston was able to give us a 2 GB kit, the KHX11000D3LLK2/2G, which is based on Elpida chips. Results are better as we were able to obtain stability in the following settings:


- DDR3-800 in 5-5-5-15 at 1.5V
- DDR3-800 in 6-6-6-18 at 1.5V
- DDR3-1067 in 5-6-5-15 at 1.9V
- DDR3-1067 at 6-6-6-18 at 1.5V
- DDR3-1067 in 7-7-7-20 at 1.5V
- DDR3-1067 in 8-8-8-22 at 1.5V
- DDR3-1333 in 6-7-6-18 at 1.9V
- DDR3-1333 in 7-7-7-20 at 1.5V
- DDR3-1333 in 8-8-8-22 at 1.5V
- DDR3-1333 in 9-9-9-25 at 1.5V


As you can see, two settings needed more than 1.5V, DDR3-1333 in 6-7-6-18 and DDR3-1067 in 5-6-5-15. However, given that these chips are certified for functioning in DDR3-1375 at 1.7V in 8-8-8-23, and also given the results of the Corsairs, even supporting these timings is already a very good thing.

DDR3 vs DDR2

DDR3 vs. DDR2

Here, we were first of all interested in power consumption on the Gigabyte card, which is able to manage two types of memory. In order to obtain precise figures, we measured the power supplied to the motherboard at its ATX plug with an ammeter. This allows us to eliminate processor consumption and part of the graphic card‘s.


In rest, there is gain of about 2 watts compared to DDR2 (2x1 GB of Corsair PC8500), and this went as high as 5 watts in load (with Prime95). It should be noted, however, that to have equivalent timings we had to increase the DDR3 to 1.8V or the same power consumption as DDR2. Here, consumption was then more than 1 watt in rest and less than 1 watt in load. For a desktop PC, DDR3 doesn’t change much in this area, however, for a laptop it could be of greater consequence.

The P35 Express chipset offers a number of bus processor and DDR3 frequency ratios. Here are the ones available on the P5K and P5K Deluxe :

- FSB1066 : DDR3-800, DDR3-889, DDR3-1067
- FSB1333 : DDR3-800, DDR3-833, DDR3-1000, DDR3-1067, DDR3-1111, DDR3-1333

Let’s now see the performances of DDR3, first with a Core 2 E6750 set at 2.67 GHz in FSB1333 on the P5K and P5K3. We used four measurements; the rate in MB/s and the latency in cycles measured by ScienceMark 2, compression time of files with WinRAR 3.7, and finally, the framerate in Far Cry. These two « practical » applications were chosen because they are noticeably affected by the speed of the memory sub-system, something that isn’t always the case.


The first thing we see is that DDR3-800 in 5-5-5 is slightly slower in terms of latency, as we expected and this places it 1 to 1.5% behind DDR2 in our practical tests. The second observation is that DDR3-1333 only has slightly superior performances than DDR2-1067.

For the latency/frequency combo, we can see that DDR3-1333 CL9 is faster than DDR3-1067 CL7, which is in turn superior to DDR3-800 CL5. In DDR2, the DDR2-1067 CL6 is equivalent to DDR2-800 CL4, while the DDR2-800 CL5 and DDR2-667 CL3 are equal. Finally, if we look at the fastest and slowest memory configurations, there is only a gain of 14% in WinRAR and 10% in Far Cry. Memory does not have as a big an impact on performances as some claim.


Does the gap between DDR2 and DDR3 change with a bigger FSB? To find out, we used a FSB1777 (444 MHz) while keeping a frequency of 2.67 GHz. Here are the results:


Increasing the FSB isn’t in favour of DDR3 and the DDR2-1067 shows an even larger gap. It seems as if the FSB:DRAM ratio of 2:3 used here is less effective than the 1:2 necessary in FSB1333, because despite the increase in bus frequency, application performances stagnate.

FSB1333 vs FSB1066, Core 2 E6850

FSB1333 vs. FSB1066, Core 2 E6850

The P35 Express is the first chipset to officially support FSB1333, however, we do have to say that this frequency is already supported by almost any decent chipset. So what does FSB1333 add in terms of performances ? Here are some tests in 2.66 GHz :


If memory bandwidth takes off with gains of around 18%, practical performances are only slightly affected at 0.8 to 1.2% in WinRAR and 1.7 to 1.9% with Far Cry.


Actually, the most interesting thing about E6x50 processors expected out in mid July aren't really their FSB, but rather their price ($266 for the E6850 at 3 GHz) as well as the characteristics of new G0 stepping. We were able to get our hands on a sample of one of these processors and they do look promising.


The first thing to notice is that the processor’s in load consumption measured with an ammeter on the ATX12V decreases 10 watts compared to the X6800. The second point is that this processor is very at ease in overclocking as we were able to attain a stable 9x375 Mhz while at the same time keeping a VCore of 1.325V. For those who are wondering, at equivalent frequencies the performances of this new stepping don’t change at all.


The only negative point is the E6850’s increase in FSB. Our sample processor was limited at around 450 MHz, which is a very low value. It’s a known fact that all Core 2s don’t have equal FSBs, but this allows us to see that an official « FSB1333 » processor doesn’t necessarily let us go higher than a « FSB1066 ».

P35 Express – Memory

P35 Express – Memory

After having reviewed DDR3, we are going to take a look at the P35 + DDR2 relationship, which will probably be the most common. Compared to the P965, as we said above, the P35 officially supports DDR2-1067. However, another characteristic is that (at least in theory) it needs a minimum of DDR2-667.

Thus, the P5K Deluxe offers the following modes depending on the FSB :

FSB1066 : DDR2-667, DDR2-800, DDR2-889, DDR2-1067
FSB1333 : DDR2-667, DDR2-800, DDR2-833, DDR2-1000, DDR2-1067, DDR2-1111, DDR2-1333

Like the P965, the P35 doesn’t allow the use of a memory frequency inferior to that of the bus system. The minimum ratio is 1:1, for example in FSB1333 (333 MHz Quad Pumped) and in DDR2-667 (333 MHz DDR). In FSB1066 the synchronous mode is no longer offered, and we must go through DDR2-667. It’s also highly probable that some motherboards simply won’t start with a memory where the SPD won’t indicate a minimum of DDR2-667 even if in practice the chip could support this frequency.

What are the performances of the P35 Express memory controller ? We measured performances of ASUSTeK’s P5K Deluxe compared to the P5B Deluxe (P965), P5W-DH Deluxe (i975x) and P5N32-E SLI (nForce 680i). First off, here are results obtained in FSB1066 :


In DDR2-800, the i975X and nForce 680i are the fastest, on the condition that the latter is in « 1T » mode. This is something not possible with all memory modules, and in the opposite case, the nForce 680i is in last place, close to the P965 and P35 in Far Cry but further behind in WinRAR. The P965 and P35 display very similar performances, which are 3-4% inferior in practice to the 975x, for example, in applications dependant on the memory subsystem. Either way this isn’t too dramatic.

These differences are made up for in DDR2-1067, because here the nForce 680i didn’t have the advantage of its 1T mode, while the i975x is too unstable with two 1 GB memory modules at this setting (it‘s officially limited at DDR2-667). Thanks to DDR2-1067 5-5-5, the P35 and P965 are at the same level as the nForce 680i and i975X in DDR2-800 4-4-4.



And performances in FSB1333 ? Here again the use of various memory settings give us very close performances in the end with differences of less than 1% in Far Cry. In WinRAR, the i975x saves itself with results 1.7% better than its closest competitor. At equivalent memories, the P965 and P35 are very similar.

While waiting for a more complete chipset comparison, we now go ahead with a duel between the P965 Express and P35 Express. Let’s take a look at some specific areas such as PCI Express management, Serial ATA, RAID, USB 2.0 and even overclocking.

P35 Express – PCI Express

P35 Express – PCI Express

On paper, the P35 doesn’t appear too much different from the P965 in terms of PCI Express management. The MCH manages 16 lines, which are only attributable to a single peripheral. This is in contrast with the 975x, which can manage 2x8 lines. The second PCI Express x16 port that we find on some motherboards actually uses the ICH’s four PCI Express lines . This of course has a non negligible impact on performances in certain cases, which we will illustrate in three separate tests :

- Serious Magic, a test which carries out a download of the graphic card from the system
- SPECViewperf9 UGS Teamcenter, a graphic OpenGL simulation, simulating a very heavy professional use
- 3DMark2006, here the score of two HDR/SM3.0 scenes in 1920*1200


The graphic card was a NVIDIA GeForce 8800 GTX , combined with an E6700 and 2 GB of DDR2-800 4-4-4-12.


The download test logically shows a strong decrease in transfer rate, which however is not divided in four. In x16, we were at 30% of the theoretical transfer rate and this figure attained 69% in x4.


SPECViewperf shows a similar performance ratio and going into x4 therefore has a large effect on performances for this type of very heavy load.


3DMark06, representing a graphic load that is closer to games of the last generation, is affected but to a lesser extent. However, the difference between a x4 and x16 port allow a performance gain from 27 to 29%!

As you can see, it is really unfortunate that Intel didn’t do better than a x4 on a second PCI Express port. The solution used for the i975x, dividing its 16 lines between two peripherals, is more advantageous, because the impact of a single x16 port on gaming performances is low.

We will have to wait for the future X38, which should allow the management of two x16 ports via its MCH, so that we can have a DDR3 chipset capable of properly managing two graphic cards, whether managed one at a time or in pairs. If it’s not SLI compatible, which remains a guarded reserve for NVIDIA and its nForce, the P35 is CrossFire compatible like the P965.

P35 Express – SATA, RAID, USB 2.0

P35 Express – Serial ATA and RAID

To test the Serial ATA interface, we used a Raptor WD1500ADFD and a RAID of two Raptor WD740GD. We will start with the performances in classic SATA mode with and without NCQ measured with IOMeter.

IOMeter is used to simulate the load in a multi-user environment, in this case by using a file server type load comprised of 80% reading and 20% writing, all 100% non sequential. In this type of operation, NCQ can be particularly useful. We tested IOMeter with a number of concurrent commands from 1 to 128. Obviously, with a single command NCQ doesn’t change anything and actually adds a small load on data managed by the chipset.


Without AHCI, the two chipsets are very close. The AHCI mode, which allows the use of NCQ is slightly to the P965‘s advantage.

We now move on to performances obtained with a RAID 0 (stripping) of two Raptor WD740GD, first with HD Tach :


Performances in reading cache as well as in sequential reading and writing are very close. The figures here are recorded maximum. Let’s see what happens with IOMeter :


This time the P35 Express is slightly faster than the P965.

USB 2.0

What about the performances of the USB 2.0 interface? We carried out a series of tests in copying 2GB of files to a LaCie External HDD and then recopied it back while measuring the transfer rate and processor use. The maximum speed of the interface is also measured by addressing the cache memory of the external hard drive via HD Tach.


Here the P35 is behind the P965 with slightly lower transfer and higher CPU use that are a little higher. Of course, this isn’t dramatic but it is noteworthy.

ASUS P5K vs. P5B – Cons. & Overclocking

ASUS P5K vs. P5B – Consumption, Overclocking

Consumption

Before we more closely address overclocking of the P35 compared to the P965 Express on two leading ASUTeK motherboards, the P5K Deluxe and P5B Deluxe, the power consumption of these two cards was measured. We used an ammeter to take readings on the different lines of the ATX block. This allows us to do away with processor consumption, which is fed by the ATX12V, and part of that which feeds graphic components on the six pin PCI Express outlet. Measurements were done with 2x1 GB of DDR2-667, an E4300 in rest and then in load with Prime95, and a Radeon X1950 Pro:


As you can see, figures are comparable.

Overclocking & FSB

Here we started with overclocking the chipset with two different CPUs, a Core 2 Duo X6800 and Core 2 Quad Q6600. We only give you the frequencies validated with two to four sessions of Prime95 for 15 minutes.

On the P5B Deluxe, we reached 510 MHz of FSB with the X6800 without touching the chipset voltage, and then we attained 530 MHz in going from 1.25 to 1.45V. On the P5K Deluxe, the 530 MHz was reached with the starting voltage, 1.4V for the chipset and 1.7V with the CPU PLL setting, which allowed going up to 540 MHz. We should also note that on the P5B Deluxe we had to go to 1.45V to reach 430 MHz in order to not have a PCI Express slot configured in x1 instead of x16. We didn’t encounter this problem with the P5K Deluxe. Like on the P965 with high FSBs, an active ventilation (or watercooling) of the chipset would be desirable and for us we did this starting at 500 MHz.


With the Q6600, results are even better because while the P5B Deluxe limited us to FSBs of 340, 360, 370 and 380 MHz with chipset voltages of 1.25, 1.45, 1.55 and 1.65V, on the P5K Deluxe we were able to obtain 470 MHz with starting voltage and 480 MHz in pushing the chipset to 1.55V and the CPU PLL to 1.7V.

Performances & FSB

Frequency isn’t everything in the domain of FSB, as the P5B Deluxe is well known for using much higher latencies on the chipset level as soon as we go to 400 MHz. Is this the same for the P5K Deluxe ? To find out we carried out various performance tests always at 2.67 GHz, but in FSB1066 (266), FSB1333 (333) and FSB1777 (444). You should note that contrary to the P5B, with the P5K the modification of the processor ratio is found in the JumperFree menu, which let’s us avoid going between two menus for this adjustment.


While P5B performances in FSB1777 are poor, frequency is not enough to compensate for latencies that are too high in relation to FSB1333. The P5K doesn’t suffer from this problem and therefore has performances that are 3 to 4% higher.

Frequency & memory

We end this test with a test in increasing memory frequencies with a FSB:DDR ratio of 1:2, carried out with 2x1 GB of PC2-6400 Crucial Ballistix Tracer configured in 5-5-5-15 at 2.25V, and then with 4x1 GB of the same type of memory.


On the P5B Deluxe maximum frequencies are 590 MHz with two memory modules and 570 MHz with 4. Increasing the chipset power did not have a positive effect on the increase in frequency.

On the P5K Deluxe, at the starting voltage we are at 600 MHz with 2 memory modules and 550 MHz with 4. The increase in chipset voltage did not allow going any further with 2 memory modules, however, with four we obtained 560, 580 and 590 MHz at 1.4, 1.55 and 1.7V, respectively.

Conclusion

Conclusion

A simple evolution of the P965 Express, the P35 Express does offer anything earth shattering. Like DDR2 during its era, DDR3 will also take some time to impose itself with the arrival of memory that have better frequency and latency combinations. In the meantime, DDR3-1333 doesn’t give us notable performance gains compared to DDR2-1067, and while the price of DDR2 is at an all time low, this necessarily more expensive memory is like a fly in the soup.


As they say the first time is always the hardest and this also applies to new memory technology. DDR3 should not distinguish itself from DDR2 before the arrival of chipsets that manage DDR3-1600 and quad core processors in FSB1333 or 1600 capable of gaining in available memory bandwidth. For this reason, combo motherboards that look good on paper aren’t necessarily that great in practice. We might as well opt for a classic model solely in DDR2 and a mode where performances are equal to the P965‘s.

Besides the memory aspect, it’s unfortunate that the P35 doesn’t have a more flexible management of MCH PCI-Express lines, as is the case with the i975x. Of course, Intel will soon offer on its X38 thirty-two PCI Express lines, allowing it to manage two ports in x16. These ports will be additionally in PCI-E 2.0, a norm which allows a doubled transfer rate as well as the possibility of going up to 150W directly by the port. In practice, this bandwidth should not be really usable for several GPU generations and power can still be supplied via external connections. For this reason, for PCI-E 2.0 graphic cards that are 1.0 compatible, the presence of PCI-E 2.0 does not seem to be an essential characteristic.

Now the only thing is that the management of 2x8 on the P35 would not have been a luxury and would have improved performances whether if cards were in CrossFire or not, SLI being the reserve of the nForce.


We finish on a positive note for the P35 Express as it was integrated on ASUSTeK’s P5K Deluxe. While for classic use this motherboard doesn’t change much compared to the P5B Deluxe, in overclocking it has non negligible improvements thanks to the chipset. The raise in frequency is even easier than on the P965, notably on the quad core, all with better performances and Penryn compatibility, a characteristic of all P35 cards. Who could ask for more?