ASUS P5E3 Deluxe Overclocking: DDR3 Takes Front Stage
by Kris Boughton on November 20, 2007 4:00 AM EST- Posted in
- CPUs
General Memory Performance Scaling and Command Rate
Fiddling with the BIOS to see what works and what does not is all well and good but meaningless unless we have A way to quantify the results. To that end, we've done our best to tabulate the details of our testing and present the following.
Before delving into more advanced discussions let's first review what we already know about typical memory scaling. In this graph, we show how memory responds to a simple change in base operating frequency. Here we can see the memory responds just as we expect - higher frequencies bring about higher bandwidths.
The green lines represent actual data points, collected with timings of 8-8-8-15, a Static Read Control Delay of 7 and Command Rate 1N. The blue lines represent extrapolated data points - i.e. we did not measure these values directly but rather scaled existing test results to create additional information for comparison purposes. This is an important detail to note as this means we extrapolated the data points in the context of our original tests (the same timings, etc.). While this isn't 100% accurate, it does provide a very good estimate for the timings you would most likely set at these particular speeds.
The result is an unadulterated illustration of memory read performance with frequency being the only variable - just as we intended. This also gives us the ability to predict how our memory would perform should we run it at DDR-2000 with 8-8-8-15 timings. While some choose to do this by pumping excessive voltage through their memory modules, we feel our method is safer as we do not risk damaging our expensive DDR3. Future DDR3 will likely be able to reach these higher performance settings at more reasonable voltages, however.
Our ASUS P5E3 Deluxe exhibited an amazing memory scaling range, allowing us to run our memory 1:1 at stock speeds with tight 4-4-4-12 timings and as high as about DDR-1960 at 8-8-8-15 by using the more exotic 2:1 divider. It appears our final limitation was the board/CPU combination and not the memory. We have no doubt the memory would have continued to scale well above 2GHz had there been additional FSB headroom to play with.
In this analysis, intended to show the affect of 1N versus 2N Command Rate, we see that setting 1N (DRAM Command Rate to 1T) can improve synthetic memory read scores by as much as ~3.3% over 2N. Static Read Control Delay was held constant at 7 through the use of Ai Transaction Booster while both DDR3 memory frequency and timings were increased and Command Rate was forced to either 1N or 2N by BIOS.
We can conclude that DDR-1860, 8-8-8-15-CR1 is roughly the equivalent of DDR-1925, 8-8-8-15-CR2. This means modules that might not otherwise be able to reach the speeds necessary to provide the desired target bandwidth can take advantage of the performance enhancing capabilities of the MCH in order achieve this end. In most instances enabling Command Rate 1N did nothing to bring about early instability making this setting nothing more than free bandwidth for the taking. The source of the increased bandwidth comes from the addressing efficiency improvements that come with 1N commanding and the reduction in overall memory access latency.
Fiddling with the BIOS to see what works and what does not is all well and good but meaningless unless we have A way to quantify the results. To that end, we've done our best to tabulate the details of our testing and present the following.
Before delving into more advanced discussions let's first review what we already know about typical memory scaling. In this graph, we show how memory responds to a simple change in base operating frequency. Here we can see the memory responds just as we expect - higher frequencies bring about higher bandwidths.
The green lines represent actual data points, collected with timings of 8-8-8-15, a Static Read Control Delay of 7 and Command Rate 1N. The blue lines represent extrapolated data points - i.e. we did not measure these values directly but rather scaled existing test results to create additional information for comparison purposes. This is an important detail to note as this means we extrapolated the data points in the context of our original tests (the same timings, etc.). While this isn't 100% accurate, it does provide a very good estimate for the timings you would most likely set at these particular speeds.
The result is an unadulterated illustration of memory read performance with frequency being the only variable - just as we intended. This also gives us the ability to predict how our memory would perform should we run it at DDR-2000 with 8-8-8-15 timings. While some choose to do this by pumping excessive voltage through their memory modules, we feel our method is safer as we do not risk damaging our expensive DDR3. Future DDR3 will likely be able to reach these higher performance settings at more reasonable voltages, however.
Our ASUS P5E3 Deluxe exhibited an amazing memory scaling range, allowing us to run our memory 1:1 at stock speeds with tight 4-4-4-12 timings and as high as about DDR-1960 at 8-8-8-15 by using the more exotic 2:1 divider. It appears our final limitation was the board/CPU combination and not the memory. We have no doubt the memory would have continued to scale well above 2GHz had there been additional FSB headroom to play with.
In this analysis, intended to show the affect of 1N versus 2N Command Rate, we see that setting 1N (DRAM Command Rate to 1T) can improve synthetic memory read scores by as much as ~3.3% over 2N. Static Read Control Delay was held constant at 7 through the use of Ai Transaction Booster while both DDR3 memory frequency and timings were increased and Command Rate was forced to either 1N or 2N by BIOS.
We can conclude that DDR-1860, 8-8-8-15-CR1 is roughly the equivalent of DDR-1925, 8-8-8-15-CR2. This means modules that might not otherwise be able to reach the speeds necessary to provide the desired target bandwidth can take advantage of the performance enhancing capabilities of the MCH in order achieve this end. In most instances enabling Command Rate 1N did nothing to bring about early instability making this setting nothing more than free bandwidth for the taking. The source of the increased bandwidth comes from the addressing efficiency improvements that come with 1N commanding and the reduction in overall memory access latency.
Facebook
Twitter
RSS
25 Comments
View All Comments
frede86 - Tuesday, September 2, 2008 - link
hey folksnice guide u made there m8.
but ive tryed to use that setup u recomment.
but doenst work. how come? is it because i use a dou core E8500?
Cheers
frede86 - Tuesday, September 2, 2008 - link
Core 2 dou*cEvin Ki - Saturday, February 23, 2008 - link
after reading the information on the AI transaction booster, and the Memset program, i decided to brave up, and give it a go. Memset indicated that my performance level was a 7. as my ddr2 CAS was 4, i assumed that the bios was relaxing my system a little. i simply disabled the booster option in bios with a relax of zero. rebooted, and re-ran Memset. nothing had changed. still a 7. any settings other than disabled and zero, in bios, will not POST. Memset allowed me to change the performance level to 6, apply, and save the change. nothing has changed in bios as a result of that change.my question is, what am i doing wrong, as i would expect to have seen something different in Memset with changing the bios to disabled and in effect lessening the relax?
i apologize if i have somehow missed the whole point, and do not understand this memory tweaking concept.
thanks
jwigi - Thursday, February 14, 2008 - link
Hi I have a P5k Premium and the contact between the heatsinks and the board aren't very good, i was wondering what size of screw you used and also if you needed to put any springs on them, i'm thinking of doing the same 'mod' you've done in your article on my board...thanks
plextor10000 - Thursday, January 10, 2008 - link
I was already one day playing with the settings of the mainboard, first tried to boost the E6850 from default 3Ghz to 3.6 , but could not make it stable in benchmarksSwitched to the Q6600 - and followed the guide , decrease the voltage for the CPU to 1.375 , for safety .
After step by step, i increased from 2.4 without any issue to 3.6 , running stable with my patriot 1333 on 1600
Thank you for this guide. Can i use the same settings for the E6850 also , or do i modyfie some settings to blaze the clocks of it ??
Ryujin - Sunday, January 6, 2008 - link
I recently got this board, and after reading this article, I really want to follow the advice therein and remove the thermal pads underneath the heatpipes/heatsinks and replace them with thermal paste, and replace all the plastic push-pins with screws/nuts.I've yet to start fiddling with the board, as I am still waiting for the CPU to arrive. A few pieces of advice I was looking for to ease my mind though:
- Would I be mad using Arctic Silver thermal compound, considering conductivity issues? (I could get ceramique, which is non-conductive, but it'll take quite a while, through the channels I wish to use).
- What diameter / length screws do you recommend? I figure 10mm M3 screws with lock nuts should do the trick... If they're too long, I should be able to screw them in with the heads facing the MB-tray.
- I'm going with a liquid cooling solution for my CPU. The P5E3 Deluxe included two fans that can be placed atop the heatsinks surrounding the CPU-socket for just such an occasion. However, I suspect they're rather noisy (are they?). Also, the case I'm using is the Coolermaster Cosmos, which does have ample chassi fans, so I'm wondering it is really necessary (time will tell, but I was wondering if anyone has any opinions on the subject).
cheers
kmmatney - Tuesday, November 20, 2007 - link
OK, looking at the graphs, it just seems like all I'm seeing is the benchmarks getting better with higher overall cpu speed. The overclocking guide was good, but the benchmarks are hard to figure out, since memory speed and cpu speed are getting higher at the same time.kmmatney - Tuesday, November 20, 2007 - link
OK, figured it out - we just need to compare the Asus P5E3 scores versus the Asus Maximums scores at (8 X 465) to see how much DDR3 improves things over DDR2. Seems to be 1%-5%. Yawn...TA152H - Tuesday, November 20, 2007 - link
When I read the article for the x48, I mentioned that it made no sense for the three chipsets unless the x48 was DDR3 only. Well, I have found out from another site that it will be DDR3 only.That makes the x38 really only useful as a DDR2 chipset, after the x48 is available. This assumes Intel did the right thing of course, and all the ugly overhead for DDR2 is removed from the x48. But if it is, you'd have to be a fool to buy the x38 with DDR3, since it is second best, and has overhead from a function that will not be present on the motherboard. It will give you more heat, and more power use for something that is completely useless. I didn't like Intel including both, but I guess it was to transition to DDR3, so it was a necessary evil until the x48 comes out.
retrospooty - Friday, November 23, 2007 - link
actually, that isnt true. X48 is just an X38 selected out of speed bin to be the fastest. They were going to market it as only DDR3 (that was a marketing decision not a functionality decision) but have since changed their minds.http://www.fudzilla.com/index.php?option=com_conte...">http://www.fudzilla.com/index.php?optio...amp;task...
Either way your arguments are pretty one sided. Even with DDR3 highly overclocked to 2ghz its really only a slight bit faster then DDR2. In fact DDR2 at 1000mhz 4-4-4 beats DDR3 at 2000mhz @ 9-9-9 in most real world tests and apps. Intel is currently going with tri-channel DDR3 on the next gen CPU (nehalem) with internal memory controller. Then and ONLY then is DDR3 going to be worthwhile, and even then its only worthwhile because Nehalem chipsets wont support DDR2. DDR3 is a minor speed bump not worthy of spending money on until Nehalem comes out.