EVGA SR-2 dual processed Xeon overclocking help.

DamoTourettes · 29 Feb 2012 at 10:31

Hi all,

New to overclocking Socket 1366 (but have loads of experience on Sandy Bridge), and am just getting my head around QPI etc. Wondering if anyone can give me some pointers on the overclock I've got below please? I was hoping for a higher OC than 4.2Ghz - I'm on air cooling however and things are getting very hot. Might some of the OC settings below help with bringing temps down/bringing frequency up?

Also, the other things I'm interested in are:-

1. "Signal" settings - can adjusting these via the IOH etc affect OC performance?
2. In having the X5690 Xeons the multi is set to 26 - my understanding is that I should leave this as high as possible and use the frequency settings from here to adjust the OC - is this correct?
3. These chips can have a max QPI speed of 6.4GT - should I leave this at default of 4.8GT or increase it?
4. Is the memory speed OK at default at 1066Mhz?
5. What's the MCH strap?
6. Last but not least, Uncore frequency? What should I be changing this to?

I appreciate these are all absolutely noob questions, however this really is a bit of a new thing for me - as such I welcome all pointers - thanks a lot in advance...

Default E-LEET page

Voltages

Temperatures (under full load running Intel Burn Test/Prime95)

Overclock settings in BIOS

Voltages1 - BIOS

Voltages 2

Signal (No idea what this even does?)

RAM

DamoTourettes · 1 Mar 2012 at 10:46

*bump*

Anyone?

Krony · 1 Mar 2012 at 12:43

Micht wanna try on the Evga forums in the Classified SR-2/SR-X section.

DamoTourettes · 1 Mar 2012 at 12:48

Tried that - no dice unfortunately

Apparently it's just like a normal i7 x58 to overclock, but I can't seem to fathom QPI voltages etc. Looking for about a 4.2Ghz OC with reasonable temperatures...

kerrgreg · 1 Mar 2012 at 13:35

I started to give you some help over on the EVGA forum.

I have not changed the QPI link voltage from default on mine and was able to run 180bclk before one of my ES chips died.

What exactly are you having problems with? what have you been able to achieve so far?

DamoTourettes · 1 Mar 2012 at 13:50

Oh hey - thanks for your help thus far. As far as I can tell, QPI link voltage is set at default...

In setting the multi to 26 and the BCLK to 162 I can get ~4.2Ghz (what I'm after), however how this correlates to the mem frequency is what I'm questioning. My understanding is that mem frequency is directly related to BLCK (in this instance 162), however I'm unsure whether this is going to be a preferential approach rather than using a multi of 21 and a BLCK of 200 - I've got 2000Mhz Dominator RAM in the box, but is pushing the QPI frequency preferable to running a higher multi with lower BCLK?

Also, where does uncore come into all of this?

Apologies for such intrinsically noob questions - I understand they're easily answered, but I'm really green on x58 overclocking...

Thanks a lot in advance.

DamoTourettes · 1 Mar 2012 at 14:23

Also, at boot, I have massive differences between both of the chips in terms of temperatures. At stock volts and proc speeds, I have ~30 degrees C for proc 1, but around ~45-50 degrees C for proc 0. :eek:

Is this something you experienced yourself? Anywhere up to 10 degrees I could understand, but 15-20 seems absurd...

kerrgreg · 1 Mar 2012 at 14:23

I will help were I can for some parts of your post I genuinely do not have the answer.

By my understanding you want to keep the BCLK down as low as possible since the board is limited to circa 200bclk and it is hard to push past 190+ also this is directly strapped to the QPI frequency so as you increase the QPI frequency you must also note that when you exceed the 6.4 that your chips are rated for you are likely to exceed the stable link speed.

I believe to set the memory frequency on these is a little harder, at least I have not understood it yet, It appears you set the strapping you want and then as you increase the BCLK from there it increases but its not clear at what ratio. I have seen lots of people are running 2000mhz ram at circa 1600 with tightened timings. this might be because the BCLK being limited to about 200 seems to stop you hitting the really high ram speeds.

Personally I would look to leave turbo on and run at 26x what ever you can achieve.

There are a number of threads on the evga forum with some voltages and settings.

I used these:
Link 1
Link 2
Link 3
Link 4

kerrgreg · 1 Mar 2012 at 14:26

By the sounds of there might be a problem with your CPU HSF what ones are you using?

There is a backplate issue with the CPU socket near the rear IO plate where something gets in the way this might be the issue I didnt know about this issue before I installed my Heatsinks and have not had an issue yet so its hard for me to tell you more about that.

DamoTourettes · 1 Mar 2012 at 14:34

Sorry, CPU HSF? Not sure about that terminology, sorry.

As for the overclocking, thanks for the pointers. I was actually able to get ~4.2Ghz (what I was after) with 26 x 162 BCLK, @ ~1.27 vcore, however I was unsure of the memory timings etc on it as well. As in the photos above I've set the memory at 1066 and understand that the BLCK will affect this, but I don't know what that equates to with regards to exact memory timings. I'd be happy to run my 2000Mhz memory at say 1600Mhz with tighter timings (default is 9-10-9-28 @ 2000Mhz), but 7-7-7-27 or something would be fine. What was the final OC you got out of your box yourself? Do you think it's worth returning the hardware (i.e. the hot chip?)

DamoTourettes · 1 Mar 2012 at 14:35

Oh, and again, thanks a LOT.

kerrgreg · 1 Mar 2012 at 14:44

CPU HSF == CPU heatsink fan.

Personally having had a bad experience with my ES CPU's I would recommend taking a slightly lighter overclock to aid a cooler running mobo in particular. Since there is a power mosfet north of CPU1 and that will run substantially hotter than the CPUs.

Sadly I was unable to push my CPUs more than 170 as i believe I fried one of my ES chips (took the risk buying them) so will be buying retail chips.

personally I would bump your ram to 1333 to match the fsb and see where it goes to since you are unlikely to hit 2000mhz and you will be running on spd then once you find the pay off between OC and heat then you can work on tightening the timings.

Hope that helps

DamoTourettes · 1 Mar 2012 at 15:02

Many thanks - in having D1 Xeon 5690's with a multi of 26 I haven't had to push them past a BLCK of 162 to get 4.2Ghz, however they're running pretty hot. Unfortunately the temp differences means one of the chips has pinged mid 80's under load whilst the other is early 70's - something I can live with. The box is going to be running MATLAB software for extended periods (i.e. days on end at 100% CPU), so I want temperatures to be totally fine - and it appears that this difference in temps is what's limiting my OC. As for the memory timings, I'm OK with that, but how does adjusting the RAM to 1333 over 1066 change things? I thought RAM speed was directly correlated to the CPU BCLK? Also, can you illuminate me as to what the MCH strap does?

Sorry for so many questions...

kerrgreg · 1 Mar 2012 at 15:28

I am not entirely sure about this and am still learning about it myself so I am going to paste this stolen from another forum about the SR-2:

Two different things there....
MCH Strap is one thing, Uncore multiplier is another.

Your uncore multiplier should technically be no higher than 2x your ram divider. IE if your ram divider is 2:8, your uncore should be 16. (since 8x2=16)
If you're running 2:10 your uncore might be 20x but you generally need a lot of QPI/Vtt voltage (depending on what the bios calls it) to run a 20x uncore. Most OC guides tell you to lower uncore to 16x because it will take uncore out of the equation on stability (and also allow you to use a lower Vtt voltage). Some people try to run higher uncore frequencies (17x is generally usually stable; 18x is hit or miss)...AFAIK uncore doesn't really affect F@h at all and running a higher uncore than 16x does take more Vtt so essentially it's not worth it.

Your ram dividers are fairly self explanatory....however in most bioses (including that of the SR-2), the dividers do not give a x:y divider ratio, they give a speed. If speed is given, you can safely assume that the speed is always based on the stock 133mhz frequency (in general most bioses all base any speed setting on 133mhz). IE in the SR-2, the ram settings are as follows:
DDR-800 (divide 800/133 and you get 6, so that is 2:6 ram divider)
DDR-1066 (1066/133=8, so 2:8)
DDR-1333 (1333/133=10, so 2:10)
I don't recall off the top of my head if the SR-2 has a DDR-1600 but obviously that would be 2:12.

The MCH Strap setting does not use the 133mhz base speed for what it says (as far as I can tell). IE setting MCH Strap to 1600 uses a generic 1600mhz speed settings for CAS latencies and such.

To answer your question - if you want to OC your ram, you would change the ram speed setting to increase the divider. The end result will be your base clock (bclk) multiplied by the 2nd number of your divider (ie 6, 8, or 10). IE in my case I'm running 200bclk, with my 1333 ram oc'd to 1600 I was using the 2:8 divider and therefore my ram speed setting was DDR-1066 which is the 2:8 divider. Now I'm running DDR3 2000, same bclk, so I changed the setting to DDR-1333 (2:10, 200x10=2000 for the ram).

Obviously depending on your ram it might work, it might require more vdimm voltage, or it might not work at all if you oc too high. You can usually get a few hundred mhz free on ram with no voltage bump, but depending on what your bclk is, changing the divider might push the ram too far. You say you're running 200bclk; if your ram is 1600 your divider should be 2:8. However you probably won't be able to change it to 2:10 and get your 1600 ram to run at 2000. If you're running 1333 @ 2:6 divider and you change it to 2:8 like I did to run @ 1600, that will probably work no problems. Going from 1333->1600 is fairly easy and a safe bet; going from 1600->2000 is a 400mhz oc for the ram and may not be possible without more vdimm or looser timings (ie higher # timings like 10 or 11 instead of 8 or 9).

Link

I would certainly say there is a problem with your cooling. either there blocks/ heatsinks are sat incorrectly on the cpus or the backplate is incorrectly installed.

Personally I would go with a lower overclock since matlab is multi threaded pure ghz should not slow you down too much.

Also what case are you using? as well as what CPU heatsinks?

It looks like a number of your voltages are quite high.
Specifically if you can get the IOH voltage down from 1.275 to 1.1 What I'm running
Same with your VTT try getting that down I'm running 1.2
run your ram at a lower voltage if its possible as that will also help get the power down and heat.

DamoTourettes · 15 Mar 2012 at 14:40

Hi Greg,

Your helps really invaluable - thanks very much for all the pointers - they're really appreciated.

Have upped mem speed to DDR-1333, so am now running a 2:10 divider, considering I've got a 26 multi and bclk set to 162, that's giving me about a 1620Mhz mem overclock, which I'm happy with. The memory I've got in the box is pretty low latency (even at 2000Mhz), so I've pulled the timings back to 7-8-7-20-2N and everything seems stable. Cheers for the pointers as well on the voltages too - have pulled back the IOH and VTT to your numbers and am running OK, and am hitting an absolute top (on one core) of 72 degrees after 4 hours of Prime95. It's still a little high, but most of the other cores average about ~65 degrees which I'm happy with.

Again, many thanks for your help.

kerrgreg · 15 Mar 2012 at 19:49

No problem

subbytna · 17 Mar 2012 at 23:31

you have 12 cores running at 4.2.....do you REALLY need more performance

DamoTourettes · 18 Mar 2012 at 01:33

...AND the 2 x EVGA GTX 580 Ultra Classified's clocked to 900Mhz on the core.

NEED. MORE. POWERRRRRRRRRRRRR...

kerrgreg · 18 Mar 2012 at 10:29

Nice work damo

DamoTourettes · 29 Mar 2012 at 17:07

Hey all... back again.

...so the machine stabilised with the settings above (brough IOH etc down, temps reduced, all good), with 24GB of RAM, but the developer here at work whose machine it was asked for another 24GB. I put the RAM in and it took it fine, however in running Prime95 now I'm getting the "worker stopped" whilst running an FFT stress test. I understand that increasing the amount of RAM in some instances will lower an overclock, however I'm wondering what I can do to get around it - my RAM is running at 1620Mhz (based on a 26 multi and and BCLK of 162), along with RAM volts of 1.60v - will increasing the RAM volts to 1.65 help this issue, or should I be increasing IOH, VTT etc to stabilise things?

Thanks (as always) in advance.