What I really need to do is go back to "first principles". Isolate the BCLK and find the maximum, isolate the memory, and isolate the processors. Adding voltage hasn't helped, and I played with the memory settings today and that didn't accomplish much either. I need to find out where the limits are so I know what's holding me back. I may just have one CPU that doesn't want to OC.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
SR-2 Optimization Thread
- Thread starter musky
- Start date
Mine will drop a bigadv in a few hours...I'l fire up LinX at 10Gb to see if it is truely stable at 198 bclk. I agree strongly with MIBW. Memory timings are tricky. 8-8-8-20 on mine works fine, but 9-9-9-24 would not work over 175 bclk.
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Yeah but I every Intel CPU I have personally overclocked since since Core 2 duo days (7 CPUs) will overclock at least 140 - 150%. Given the hell I have gone through for weeks with memory problems (OCZ and Kingston) I find it highly unlikely the CPU is the limiting factor and crapping out at 130% overclock.
The trouble with the first principles method is working out what is really maximum? You can keep cranking the volts and keep going, but I worry a bit about long term frying things. My reading of the forums made me settle on 1.4v core, 1.375v vtt and 1.4v IOH as the maximum I was prepared to run 24/7 on air cooling, so my efforts have been a quest to see what I could get for that. Luckily the answer is 4 to 4.2 Ghz. Usually I am temperature limited in my OC, but this is the first time I have had cooling up to the task, and so I am really voltage limited. I know I could get 4.4GHz or so at 1.45v, but that is not an option for 24/7 use, esp on air. I will probably drop to 4.0 in summer, as 1.31v 1.375vtt and 1.375 IOH is very manageable with the Noctua DH-14s.
Anyway, I won't push if you do not want to share settings, but being self taught and by no means an expert at this I love seeing other peoples settings, as a chance to learn something. Shaminos template got me up and running very quickly, and saved a great deal of time.
And like Musky - I just finished a bigadv (my first in weeks), and you have inspired me to run linpack again at max mem usage, just to be sure
Last edited:
My current settings haven't changed since earlier in the thread, except that ACPI T state is disabled (I think it was before but it wasn't mentioned) and MCH and Uncore are auto. I tried your BIOS settings exactly, tried changing MCH, tried adding voltage, tried command rate 2T, etc. If you guys can pass LinX then I'll know it's just me. I agree on your voltage limits too, so I don't plan on testing beyond those.
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
I have no idea if ACPI T state makes a difference - it did not hurt my earlier efforts, but I saw it off in Shaminos guide so off it goes.
So you have tried a loose manual set of timings, known to work at lower clocks, then set MCH 1600 and uncore 12x and DDR800 2:6? and no pass?
The reason I ask is that my first set of OCZ memory (sent back) was unstable at 2:8 at any bclck, but rock solid at 2:6.
My second set of Kingston passes everything at 2:8, with just a slight loosening needed for 12 dimms from 7 8 7 20 1T to 8 9 8 24 1T etc.
But, the thing that is really bad about the SR2 is inconsistent booting. Before loosening, the 7 8 7 20 1T settings would pass everything for 12 hrs etc.... on about 9 out of 10 boots. Sometimes it would boot and be completely unstable. Fail anything in 5 mins. Reboot and pass 12 hours.
Nightmare.
It takes quite a while on mine, like half an hour a frame at 20GB, which should mean about 15 mins a frame for 10GB - but I have no idea if IntelBurnTest2.5 is running the same flavour of linX as you.. shame it failed, might need to loosen some settings.
Here is my pass at 8 hrs and still going, just so you don't think I am a purveyor of bull-merde..
http://www.pbase.com/davexl/image/128370459/original.jpg
Mine only goes about 5 minutes before it crashes (the LinX app crashes, not the machine.) I guess since this one is not a daily driver and so long as it is bigadv-stable, I am fine with it. I don't doubt I have a memory module that is less than optimal.
Thanks for testing, musky. I see the exact same thing. Each pass takes about 5 minutes, and with a higher overclock it will say that it failed after one pass. Maybe the LinX app is just extra sensitive? I'll have to try IntelBurnTest for comparison.
It does make a difference, but it's very slight I think. After upgrading the BIOS it was on by default, and I couldn't even push BCLK from 175 to 176. I turned it off and 176 seemed to work.
I haven't tried 2:6 yet, but I'm running the Corsair at the specified 8-8-8-24 settings. This is why I want to isolate and test the memory and CPU separately to see what's holding me back. I don't know when I'll have time to do that, though.
I saw this too. My "auto" memory settings had one value at MCH Auto and manual 1066, then they went up when I set MCH to 1333, but they didn't come back down when I set MCH back to 1066, even with a power cycle instead of a reboot. I don't know if disconnecting power would help or not.
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Like my screenshot above showed temps at maximum ram, 32 threads, this post shows temps during different tests, which I found interesting, but then I am not completely normal.
I am used to Prime95, where smaller FFTs usually result in higher temps. Why a lower temp here I hear you say? Argus Monitor only records CPU0.
Auto threads is interesting. It still seems very effective for discovering faults. It uses between 50 and 100% of CPU. But when it is running at the 50% phase of its frame, the really interesting thing is to switch taskmanager to one graph per NUMA mode. This shows that it favours the cooler running (co-incidence?) physical CPU1 over CPU0.
Full size pic for old eyes:http://www.pbase.com/davexl/image/128380087/original.jpg
I was surprised to see physical CPU using ~ 25% and the other ~75%. Perhaps not so surprising if you imagine the 4096mb might be running mostly on one set of RAM, and thus deemed faster to flog that physical CPU rather than shunt the data over to the other.
But it always favours CPU-1. Which is good given that SR-2s usually run a lot hotter on CPU0
But all this is theoretical. In practice I have found Very High (4096MB) Auto and/or 16-32 threads the quickest to run, and the fastest to find faults. But if you fail any combo, at any setting, you are not "workstation stable". You might be stable enough for a dedicated folding box however.
I am used to Prime95, where smaller FFTs usually result in higher temps. Why a lower temp here I hear you say? Argus Monitor only records CPU0.
Auto threads is interesting. It still seems very effective for discovering faults. It uses between 50 and 100% of CPU. But when it is running at the 50% phase of its frame, the really interesting thing is to switch taskmanager to one graph per NUMA mode. This shows that it favours the cooler running (co-incidence?) physical CPU1 over CPU0.
Full size pic for old eyes:http://www.pbase.com/davexl/image/128380087/original.jpg
I was surprised to see physical CPU using ~ 25% and the other ~75%. Perhaps not so surprising if you imagine the 4096mb might be running mostly on one set of RAM, and thus deemed faster to flog that physical CPU rather than shunt the data over to the other.
But it always favours CPU-1. Which is good given that SR-2s usually run a lot hotter on CPU0
But all this is theoretical. In practice I have found Very High (4096MB) Auto and/or 16-32 threads the quickest to run, and the fastest to find faults. But if you fail any combo, at any setting, you are not "workstation stable". You might be stable enough for a dedicated folding box however.
Last edited:
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Interesting! I am certainly glad it is not just me going through this, but I wish it wasn't happening to either of us!
I makes testing SO hard when you can not be confident that one set of settings = same result.
I did get the feeling things were more "honest" if I pressed the reset CMOS button and loaded them again.
I wonder if this means I should look at some failures "as failure to set properly" and try again. God, I am too tired to think about that. I just want to fold and enjoy now...
I just want to fold and enjoy now...
I'm there too. I'm hoping to get my case soon, because I really need to replace my old Windows XP desktop and its mysterious BSODs. I'm pretty happy with my performance now, since I can maintain 200K PPD even with P2684s. Once I start downsizing in a few weeks I'll revisit things.
nitrobass24
[H]ard|DCer of the Month - December 2009
- Joined
- Apr 7, 2006
- Messages
- 10,466
Just curious what is your QPI frequency setting on?
I know when i was OC one of my 920s w/1333 CL9 gksill. If i left it on Auto i couldnt get above about 170 BLCK because the QPI frequency would jump up and it wasnt getting enough voltage. So i just forced it to stay at 4.8 and it was fine.
I know when i was OC one of my 920s w/1333 CL9 gksill. If i left it on Auto i couldnt get above about 170 BLCK because the QPI frequency would jump up and it wasnt getting enough voltage. So i just forced it to stay at 4.8 and it was fine.
I am starting to think that Gulftowns in general do not like high bclks, or at least they are less forgiving than Gainstowns. I had my Gainstowns to 208 bclk, Kendrak has his at 210, and Vaulter is pushing 210. All of us got there without a lot of work. MIBW and Sazan both have hours and hours trying to get everything stable and are still fighting with their systems. I spent probably 3 or 4 hours today to get from 198 to 200 bclk. It is just tough going with dual hexes from what we have seen so far...
Vaulter98c
[H]ard|DCer of the Month - October 2009
- Joined
- May 21, 2008
- Messages
- 5,817
Good deduction there, was thinking the same thing myself, but even with a limit at or above 190 the hex's still kick ass
Posted via [H] Mobile Device
Posted via [H] Mobile Device
nitrobass24
[H]ard|DCer of the Month - December 2009
- Joined
- Apr 7, 2006
- Messages
- 10,466
My Hex's will be here monday, so ill see if i have the same issue.
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Yes, I think having 50% more cores to deal with is going to stress anything to do with handling memory so much more. A bit sobering to think any oct-core upgrade in the future will likely need a downclock.
But... as for my hexes.... I think I love them. There - I had never said it out loud til just now. I am over my commitment issues to the Q6600s, it's SR2 always and forever.
But... as for my hexes.... I think I love them. There - I had never said it out loud til just now. I am over my commitment issues to the Q6600s, it's SR2 always and forever.
DooKey
[H]F Junkie
- Joined
- Apr 25, 2001
- Messages
- 13,577
My L5640 and 980x both easily handle 200 blck and higher. If you can't go above that it's a motherboard problem.
I think is has more to do with the board and all the stuff that has to work together here. I'm sure the chips are not the limiting factor, but getting them to work well with everything else in an SR-2 is tricky.
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Really? SPD is 59! I know nothing about RAM other than what I have discovered myself the hard way. It is Kingston KHX1600C7D3K3/6GX.
It passed 12 hours at 7,7,7,20,8,59,4,4,4,20,1t,0,62,65,67, but then later on I found it was not booting consistently stable. I simply recorded what was set via auto settings, then started loosening things in rough proportion. I did find that 1T / 2T had a 3% impact on folding performance, and the other timings less dramatic, so I worked to maintain 1T by loosening others.
currently at 8 9 8 24 10 67 5 5 5 24 1T auto 73 74 75.
Some RAM timings seem to have very little performance effect, so you wonder why if they hurt stability they are not set looser to begin with....
But any expert advice on what is good to loosen for stability without hurting folding performance is very welcome.
Last edited:
All this talk about RAM timings inspired me to manually set my Corsair 1600s. Guess what, it helped me eek out an additional 5mhz bclk (stable so far).
So, I'm up to 195mhz bclk for a resulting freq of 4095mhz on my E5640s
So, I'm up to 195mhz bclk for a resulting freq of 4095mhz on my E5640s
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
woot! Never say no to another 100Mhz I say. ..Was that loosening the timings? or just the benefit from manually specifying them?
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
So what you lose in timings you hopefully gain more in bandwidth and CPU MHz...But now you have me thinking if I could squeeze some more MHz out with my looser timings... damn you, I just spent 3 weeks on this, and only 3 days back folding
I captured a bunch of bigadv units so I could rerun a few frames as benchmarks - still amazed at how sensitive it is to mem timings. I am used to 3D renders, far less sensitive.
Still, I like to hear about people finding another magic 100Mhz. I had a dream the other night I found another 200Mhz with some fiddling, but then I woke up. Hmm, I need to get out more.
nitrobass24
[H]ard|DCer of the Month - December 2009
- Joined
- Apr 7, 2006
- Messages
- 10,466
At the risk of sounding stupid, how do i change the backplate on this thing?
nitrobass24
[H]ard|DCer of the Month - December 2009
- Joined
- Apr 7, 2006
- Messages
- 10,466
EFFFF
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Ahh well. Anything over the magic 4.0 is still gold. Fingers crossed for you.
Yeah, I wasted 2 hours tearing the house down looking for a 4mm allen key, luckily a neighbour had one, but sheesh, it was like postponing christmas.
nitrobass24
[H]ard|DCer of the Month - December 2009
- Joined
- Apr 7, 2006
- Messages
- 10,466
Well i went to wal-mart and got one.
Swapped the plates, but these damn brackets that came on my V6GTs suck the hit the MOFSET and chipet heatsink.
I have seen several SR-2s with these mounted. WTF. Now i have to go buy a dremel or overnight some NH-D14's.
@MIBW do your noctuas fit without modifcation?
Swapped the plates, but these damn brackets that came on my V6GTs suck the hit the MOFSET and chipet heatsink.
I have seen several SR-2s with these mounted. WTF. Now i have to go buy a dremel or overnight some NH-D14's.
@MIBW do your noctuas fit without modifcation?
may i be worthy
[H]ard|DCer of the Month - December 2010
- Joined
- Aug 17, 2010
- Messages
- 1,186
Sorta - here is what I told a guy on the EVGA SR2 forum:
I use Noctua NH-D14 CPU Coolers. I hesitate to recommend them, because there are issues. Didnt actually have to mod anything, but you need a 3 or 4mm allen key to take off CPU backplate - that stopped my build dead for a few hours!
The clearance over RAM is minimal - so no heatspreaders. Bugger to change ram with the heatsinks installed, but it can be done. They are so big they can only be mounted blowing bottom to top of case, not front to back. Suits the Lian Li PC-P80B.
They are very heavy and large, so far I have not seen any warping or bending problems, but it should be kept in mind. The lower fan would just touch the top of something in the top PCIE slot. I run my GPU right down the bottom 16x slot - which needs the optional power connector to power the lower slots.
Tall enough to lose the Lian Li PC-P80B bracket to secure multiple heavy GPUs, but that is a trivial thing to make.
Those are the negatives. The cooling is great. Twin fans = powerful, and redundancy. Well made, great fan clips.
Folding temps are in the sixties on CPU-1, and low seventies on CPU-0, at 20c ambient. 1.39vcore on x5650 hexacore at 4.2ghz.
Only other thing I did was add a few extra 120mm fans in the empty drive bays to keep pushing along cold intake air to the intake of CPU-0, and as general cooling for the region around CPU-0 socket. I taped shut air gaps in the front door between the 3 140mm door fans, to help them suck. I reversed the rear 120mm exhaust fan to be another intake, specifically to provide cool intake air for the PSU above it, which makes it run at 800-900rpm, rather than a noisy 1100+. I removed the top cover plate over the top 140mm exhaust fan, as the vents were simply way too small to get hot air out. Spoils the looks, but I don't care about that. Hmmm, maybe I should just attack it with an angle grinder.
Looks like rubbish, performs like pre-divorce Tiger.
For you, some pics (before I tidied cables a bit!)
http://www.pbase.com/davexl/image/128453819/original.jpg
DooKey
[H]F Junkie
- Joined
- Apr 25, 2001
- Messages
- 13,577
You need a special bios. Musky can send it to ya.
Well, maybe...I already had the custom bios installed, so I know it worked. I don't want to switch back because I actually don't have a chip in house that I know works with the original bios. I'd love to know if the default bios works. Looking at the patch codes in the bios, I would guess that it will not.
Kendrak, before you pull your E5530s, could you get me the CPU ID from Core Temp or HWiNFO? I have a thread on it that is buried now.
Edit: http://hardforum.com/showthread.php?t=1545163
Last edited:
that is the sexiest mess i have ever seen!
http://www.pbase.com/davexl/image/128453819/original.jpg[/QUOTE]
http://www.pbase.com/davexl/image/128453819/original.jpg[/QUOTE]
So... I am getting a couple CPUs from musky and will be putting together a SR-2. I'm wondering what I should put on my wishlist for a PSU? It will become my main rig for browsing and light gaming. However, 90% of the time it will be folding.
2 x L5640's, 12GB, dual 460's, 1 HDD, 1 SSD
PSU Recommendations? Would a good 1000W unit suffice?
2 x L5640's, 12GB, dual 460's, 1 HDD, 1 SSD
PSU Recommendations? Would a good 1000W unit suffice?