Lately there seems to be some rumble about the DDR3 CL8 world record, so I thought I am giving this one a try too. Fighting on my side are 2x1 GB Corsair Dominator GTX6 rated at 2625 MHz, 9-11-10-30, a pretty nice i7 870 ES better than aIl of the retail chips 860/870 tested so far and Besis memory freezer, which is definitely the reason why this could be working. Regarding the mainboard, I tried some boards in the last couple of weeks, but am still happy with my P55A-UD7. Does bclocks over 265+ and is quite comfortable. H55M-UD2H is too much pain in the ass.
I started with CL8, same timings as marmott and Benji Tshi used on their 3068 MHz CL8 with Kingston T1 Go and went upwards from 200 bclock. Maximum bootable setting was 250 MHz bclock (ddr3 multi 12 = 3000 MHz), but temperature had to be around -90 and more °C for the memory and between -70 and -75 °C for the cpu (very close to its coldbug). From there it was easier to do the rest with SetFSB in little steps add 1 MHz, wait a few seconds VTT was 1.56 volts, VDIMM was 1.88, but I guess it could have been lower. I got to:
World record CL8: 3078.2 MHz, CL8-11-8-31, 1T, tRFC 140, B2B CAS Delay disabled, Dual-Channel
Link: http://valid.canardpc.com/show_oc.php?id=1374351
I was also able to run SuperPI 1M at 3059.2 MHz! Other benchmark wont be a problem too should have tried MaxxMem with a bit more frequency on the cpu. :p
Well, that wasnt too hard to get, so I figured, that I should try to hit the bclock wall with CL10 with this precious 870. I tried whatever the IMC would love to throw overboard: Single-Channel, CL10-15-15-31 also played around with BCAS delay, but TaPaKaH/Sams result is unreachable (this time!). But I refused to give 1,7 volts VTT, like he did 1,62 was max seems to make no difference on my chip.
DDR3 memory clock, 2nd place: 3097.6 MHz, CL10-15-15-31, 2T, tRFC 140, B2B CAS Delay 30 (or so), Single-Channel
Link: http://valid.canardpc.com/show_oc.php?id=1334974
What should be mentioned is, that this result did not need any ln2 on the sticks (although I tried it as well). So I am still guessing pretty hard, that my cpu is limiting the results. Okay, CL10 got higher, but it seems like the IMC was just more relaxed by the loose timings and therefore allowed a slightly higher base clock. So next on my list was CL7 to really find out whats inside these heavily binned powerchips. I never tried CL7 before on those, so my expectations werent that high. I used the same settings as with CL8, just to know where Im at. I had to play around with the voltages a bit more 1.6 VTT, 1.94 VDIMM, 1.42 PCH and got it working to boot with 2880 (245*12) MHz. Now the sticks could really use the low temperature, so I had to go to about -115 °C. cpu was still near coldbug on -75 °C. I was pretty surprised when I had 3000 MHz CL7 running, SuperPI 1M stable. Here is what I finally got:
World record CL7: 3059.4 MHz, CL7-11-8-31, 1T, tRFC 140, B2B CAS Delay disabled, Dual-Channel
Link: http://valid.canardpc.com/show_oc.php?id=1374368
And a nice 1M run with 3000 CL7, whatever thats really necessary for.
For now its just a couple of useless frequencies and SuperPI times, but I will try to find some benchmarks, that can profit from these high bandwidths + low latencies. I guess I should crank up the cpu frequency too, maybe MaxxMem likes it, though we are on dual channel. Tweaking the subtimings or CL6 is also on my list hell, theres a lot to do, so stay tuned!
Here is the original thread on overclockers.at. Its in German, but adds a few images and infos: http://www.overclockers.at/extreme_c...fiziell_218663
Update 1, 9/9/2010
Hey guys, I just finished some tests with about 25 liters of ln2, so I have some results ready. But I still have to sort things out and for sure forgot a few comparisons, that would make things sweeter. Well, I need a few days more, but meanwhile I got for you the first part of my experiments:
Everest
CPU: 4900 MHz, memory: 2940 MHz CL8-11-9-24-90 1T
CPU: 5000 MHz, memory: 3000 MHz CL8-11-9-24-90 1T
Scales pretty well I guess. As I said, I didn't touch B2B Cas Delay, as it made no difference at all. Windows has a few issues with it, little bit like the R3E "slow mode" on the qpi setting.
MaxxMem @ 4.8 GHz
GTX2 @ Air, 2000 MHz CL7-7-7-18-74-1T, Score: 1466
GTX6 @ LN2, 2880 MHz CL9-11-10-30-100-1T, Score: 1797
GTX6 @ LN2, 2880 MHz CL8-11-10-30-100-1T, Score: 1837
GTX6 @ LN2, 2880 MHz CL8-11-9-24-90-1T, Score: 1859
MaxxMem likes high memory frequencies, so these results did not suprise me. Tweaking the subtimings had some effect too ... better be like that, because it's sooo time consuming. Please consider that the GTX2 was on air and not maxed out for frequency, but rather efficiency.
SuperPI 1M @ 5 GHz
GTX2 @ Air, 2092 MHz CL7-7-7-18-74 1T, time: 8.235 s
GTX6 @ LN2, 2092 MHz CL7-9-8-24-90 1T, time: 8.235 s
GTX6 @ LN2, 3012 MHz CL8-11-9-28-100 1T, time: 8.219 s
Oookay ... not much to see here. 1M likes memory, altough it's not most important. But at these rates it just does not matter. But I guess 32M would scale much more. Ran out of the precious stuff, so yes ... next time!
What is pretty interesting, is the fact, that timings can decrease performance. It seems like if one module is too weak, than the result goes back to single channel (even though it's shown otherwise in cpu-z). For example 1M at 2400 MHz with timings at CL7-8-8-24-74 1T was slower than CL9-10-9-30-134 1T. Altough CL7 is only possible under -90 °C and less.
SuperPI 1M @ CPU: 4.8 GHz, memory: 2400 MHz
GTX6 @ LN2, CL7-8-8-24 1T, time: 8.656 s
GTX6 @ Air, CL9-10-9-30-134 1T, time: 8.594
I also had the chance to shoot some new photos. Please view them here: http://www.overclockers.at/extreme_c...23#post3062423
More to come very soon!
Reply With Quote
Originally Posted by FUGGER
Bookmarks