It's all about the motherboard
You probably haven't noticed, but this review is a couple of weeks overdue. It is late because of several reasons, but foremost reason is the motherboard. I've been contemplating for weeks now what I should write in this paragraph, and to be honest, I feel that no text will do my efforts justice. When we started to plan our DDR3 review, first question that came to mind was which motherboard to use. As my fellow reviewer Keith is already working on the X48 platform, I decided to go another route. My first option was the latest nVidia chipset, the 790i. Just as I wanted to purchase this board, Asus came with its more mainstream P45 chipset and the P5Q3 board. Let's be honest here, do we really want to spend $350+ when a $250 solution is readily available (and very promising) ? You don't, and I don't want to either, because in the end, I'm writing this for your benefit, and the P45 is quite simply a much more accessible option for the most of us. But above all, while the "3" series (like P35, X38) of motherboards were not the greatest DDR3 performers, we are told that the "4" series (like P45, X48) will work together with DDR3 much better...
So the Asus P45 it is. The P45 is looking very promising indeed: around the web several people are reporting great overclocking abilities from these boards, especially front side bus-wise. 600FSB is no exception any more, so how great would it be to couple this with the excessive bandwidth offered by 2Ghz of DDR3. Out came the Asus P5Q3, and right there in its specs it says 2000Mhz DDR3 (OC). The OC between brackets is important as no cpu operates at 500FSB out of the box (yet): to be able to reach 2Ghz memory speed we have to overclock our cpu to 500Mhz front side bus.
The Asus P5Q3 deluxe, image courtesy of Asus Inc.
This seamlessly brings us to the choice of cpu : I decided to go for the E8400, a 333FSB Core 2 Duo cpu which is clocked at 3000Mhz (or 3Ghz) with a maximum multiplier of 9x. This cpu fits our purpose nicely: when we overclock the cpu to 500FSB, on a lower multiplier of 6, we're still neatly at 3Ghz. The same goes for 375x8, and 400x7,5. This way we'll be able to clock our memory to different speeds, thus providing our readers with performance numbers for different memory speeds and cas ratio's, at the same cpu speed. This is going to be a piece of cake, right ?Wrong !
When I say we're a couple of weeks late, there are some reasons of course. Things started out just fine however. The memory worked straight out of the box, and we were quickly at 1600Mhz memory speed at 400 FSB (1:2 divider). The problems started beyond 450FSB though: while the motherboard was perfectly capable of reaching 500FSB (we got up to 600FSB rather stable, but that is completely beyond the scope of this review), we ran into several issues. First of all there's the memory divider. While 500FSB with a 1:1 or 5:8 divider was a piece of cake, a 1:2 or 2:3 divider were not so easy. We needed ample voltage on the northbridge to get started, and ran into our second issue on the way: the cpu divider. It seems that, at least on the P5Q3, not all dividers are made equal. This was a known issue on the P965 chipset, but I thought we were beyond that now... As a matter of fact, the "6x" divider proved extremely troublesome, with the craved 500x6 speed difficult to achieve.
After some tweaking we got there : 500x6 speed, 1:2 divider, neatly 2Ghz on the ram. To get there though we needed to relax the performance level (PL) to a higher level than I would have liked, we needed far too much voltage on the northbridge (1.5V+), needed to tweak the VTT-voltage of the cpu, the GTL-ref values of northbridge and cpu, tweaked the PLL setting of the cpu, and played around with the memory straps to the northbridge. In short : this was not really an "out-of-the-box" experience. Have a look at the Orthos screen below, the fruit of all this effort...
Orthos screen @ 500FSBx6
This is where things went really wrong though. Although our 2D settings and tests proved to be absolutely stable, the 3D tests were a disaster. A quick run of Aquamark and even PC Mark 05 was no problem, but 3D-mark 2005 or 2006 was a no go. Whatever we tried, a crash was never far away. We can blame this on ourselves for not testing 3D-mark earlier, but whatever way you look at it, I had to go back to the drawing board.
To make a much longer story short, in the end there was no way to get 6x500FSB stable with a 1:2 memory divider. However, as soon as we tried the 8x divider things were quickly starting to clear : we could lower our Northbridge voltage, forget about NB gtl-ref voltage tuning, lower the performance level (to increase performance), lower the VTT and PLL voltages, without even stretching the cpu vcore into dangerous territory (while the cpu is at 4 Ghz now). Have a look at the Orthos shot below, and let me assure you, Crysis or 3D-Mark is not a problem either...
Orthos screens @ 2Ghz and 1600Mhz, 3D06 and Cinebench screen
So, in the end, which settings did we test ?9x333Mhz stock, out of the box settings (except for vNB), resulting in 3Ghz cpu speed, and 1333Mhz DDR(3) speed. Automatic latencies were 8-8-8-24 with a performance level of 6.
9x333Mhz with improved latencies. This means 3Ghz on the CPU and 1333Mhz DDR(3) again, at the tightest latencies we could take them, 6-6-5-22 at a performance level of 5.
8x375Mhz. Again 3Ghz on the CPU but this time 1500Mhz DDR(3), at the lowest possible latencies: 7-6-6-22 at performance level 7.
7,5x400Mhz. 3Ghz once more, and 1600Mhz DDR(3). Latencies 7-7-6-24 with performance level 8 now.
8x500Mhz. 4Ghz on the cpu now (as 6x500 did not work). The 1:2 divider did work, so 2000Mhz DDR(3) now, with 8-8-7-28 latencies, and a performance level of 11.
Finally, 8x500Mhz again (4Ghz CPU), with a 5:8 multiplier: 1600Mhz DDR(3) again with 7-7-6-24 latencies, and a performance level of 11.
We tested 6 settings thus, 4 at 3Ghz, and 2 at 4Ghz. To conclude our methodology chapter, I'd like to quote my friend and mentor Tony "Bigtoe" Leach (some might know him as OCZTony), who actually gave up on the Asus P5Q3 board himself because of its abysmal 4Gb memory performance, but that's another story entirely. Anyway, a quote, not in writing, but told to me on several occasions:
"Remember, when you are testing memory products, you are not only testing the memory, but essentially the motherboard capabilities"
|CPU || Intel Q6600 (G0) Core 2 Quad |
Cooled by Alphacool NexXxos XP Bold Highflow CPU waterblock
|Mainboard || Asus P5Q3-Deluxe P45 DDR3 |
|Memory ||2x 1Gb OCZ PC16000 Flex II @ various settings|
NEC 3540 DVD-RW
Western Digital WD740 Raptor
OCZ GameXstream 850 Watt PSU
For this review the test setup is thus based around the mainstream P45 Intel chipset. Currently you can find some very attractively priced motherboards based around this chipset, starting from around $100, but they'll be DDR2 based for that price. Overclocking wise you should get very nice performance out of the P45 motherboards, especially FSB-wise, like we mentioned before.
As soon as we ensure full stability of our memory modules on the applied settings, we can get to our test suite. We included a limited test suite more directed towards applications this time, instead of games or generic benchmarks:
SiSoftware's S.A.N.D.R.A. 2007
Lavalys Everest 4.00
Futuremark 3D Mark '06 - total score and cpu score
Futuremark PCMark 2005 - total score
Superpi 1.5, Xtremesystems edition
Crytek/Electronic Art's Crysis, through Madboris' Benchmark Tool
Techarp's H.264 benchmark
Maxon's Cinebench 10
Let's test and compare now...>