Overclocking CPU/GPU/Memory Stability Testing Guidelines

[evilqueens]

Hello there,

I'm new here on the forums (and new to overclocking too as I just built a new PC a few weeks ago) and I'd just like to say thank you very much to the OP and other members for this great cache of useful information.

I have a question question regarding overclocking the following:

CPU: Intel Core i7-3770K
Motherboard: ASUS Maximus V Formula
RAM: G.Skill Ripjaws X 16GB (4x4GB) DDR3-1600 CL9 (9-9-9-24 CR2) 1.5V
GPU: Gigabyte GeForce GTX 680 OC

So, currently, all I've done is manually tightened the timings for my RAM to the stock values (DDR3-1600 and 9-9-9-24 CR2) as the SPD detects with slightly looser timings.

I've gone on to manually set the CPU multipliers to the maximum turbo ratio (39X = 3.9GHz @ 100MHz BCLK) and I've noticed that the stock voltages under manual mode (default) can go quite high under load (around 1.2V). Therefore, I tried using offset mode and setting a negative offset of up to 0.75V and so far things have been stable. The maximum Vcore under load is now around 1.08V and running a short test of AIDA64 gives decent temperatures considering I'm still using a stock cooler (the CPU package temperature hits 74ºC max).

My question is this: I've seen some people get stock speeds at lower voltages (the AnandTech article shows a stable stock frequency as low as 0.9V). Does it make sense to set a manual voltage of 0.9V and test stability? Or should I continue increasing my offset voltage until I end up without a stable boot?

I have a Noctua NH-D14 sitting in a box - I'm just trying to establish some baseline temperatures/voltages under stock conditions so that I have something to compare to.

Thanks for your help in advance!

 

[Idontcare]

My question is this: I've seen some people get stock speeds at lower voltages (the AnandTech article shows a stable stock frequency as low as 0.9V). Does it make sense to set a manual voltage of 0.9V and test stability? Or should I continue increasing my offset voltage until I end up without a stable boot?

I have a Noctua NH-D14 sitting in a box - I'm just trying to establish some baseline temperatures/voltages under stock conditions so that I have something to compare to.

Thanks for your help in advance!

Here's what I was able to hit with my 3770k's stock HSF using stock TIM:

The problem I ran into using the offset method was that the computer would be stable under load but as soon as it went into idle the voltage fell too low and the computer would crash at idle.

So I use the manual voltage setting method when that happens. It results in higher idle volts but that is the tradeoff in minimizing the load volts.

 

[evilqueens]

The problem I ran into using the offset method was that the computer would be stable under load but as soon as it went into idle the voltage fell too low and the computer would crash at idle.

So I use the manual voltage setting method when that happens. It results in higher idle volts but that is the tradeoff in minimizing the load volts.

Yes, I was afraid of this scenario. I guess the load/idle voltages don't scale linearly (i.e a stable offset for load Vcore is not a stable offset for idle Vcore).

I see your temperatures are fairly high - even with a multiplier of 3.9X @ 0.998 V. Does this mean that AIDA64 is not a good stability tester for the CPU? Also, do you have any thoughts on PovRay and OCCT?

Lastly, would you have any numbers (similar to the ones you posted above) showing the breakdown with your current cooling solution?

Cheers!

 

[Idontcare]

I see your temperatures are fairly high - even with a multiplier of 3.9X @ 0.998 V. Does this mean that AIDA64 is not a good stability tester for the CPU? Also, do you have any thoughts on PovRay and OCCT?

Yeah you want to use LinX for testing your stability under max load conditions, there is a link for it in the OP of this thread.

The other programs are good, but LinX seems to be the best in terms of really ferreting out the instabilities in short order because it drives the temperatures up all the higher.

Lastly, would you have any numbers (similar to the ones you posted above) showing the breakdown with your current cooling solution?

Certainly Mind you my numbers are from delidding my 3770k and then using my NH-D14 or H100 (they are pretty close to the same for temperatures).

 

[evilqueens]

Wow, those are some pretty stupendous numbers. I'm very tempted de-lid at some point, but my build is still new - so I think I'll hold off for now.

I guess the temperatures without delidding/lapping would be about 10-20ºC higher?

I suppose a custom water-cooling loop would be an alternative to all of this hassle - but it just makes you wonder why Intel applied some poor-quality TIM inbetween the die and the IHS to begin with.

 

[Idontcare]

Wow, those are some pretty stupendous numbers. I'm very tempted de-lid at some point, but my build is still new - so I think I'll hold off for now.

Understandable, its a $330 gamble that you might kill your chip. Haven't read of anyone killing theirs yet though so maybe it is hard to mess it up.

I guess the temperatures without delidding/lapping would be about 10-20ºC higher?

Pretty much.

I suppose a custom water-cooling loop would be an alternative to all of this hassle - but it just makes you wonder why Intel applied some poor-quality TIM inbetween the die and the IHS to begin with.

Its not the TIM, the quality of the TIM is on par with that of NT-H1. The problem is the gap between the CPU silicon and the underside of the IHS, it is way to large of a gap resulting in too much CPU TIM coming between the two.

Checkout this post.

 

[evilqueens]

Oh I see - that clarifies things. Strange that Intel would allow this either way.

Yeah, I just ran some numbers with Linx - and my stock's are pretty much matching yours. I don't think I can get down to as low of a voltage as you can though.

Thanks again Idontcare!

 

[Kalessian]

Don't really think I need help, just sharing my experience. So I have the ph2 in my sig, it's a weird one. I had a rock solid setting but the chip is getting older and I started crashing again playing QLive. Weird part is, I add volts to the chip and Prime95 reports errors faster. I can't tell if it's heat related or the PSU/mobo can't deliver the increased volts with stability or what. It doesn't ever go above ~56C under load, usually at ~52C (the per-core temps are unavailable since I've unlocked cores). I get Prime95 Large FFT errors at ~10 hours at 1.45 and in 1-3 hours at 1.46 VCore in CPUZ.

I have the CPU-NB at 2600mhz and I'm thinking the instability is coming from there. Testing this now.

So another weird thing is whenever I crash in a game, the PC becomes very unstable even after the load has been removed and only a reboot will fix it. I can't even start any new programs as they crash instantly, all open browsers crash, smaller progams continue to run fine. Happens every time a game crashes to desktop, but doesn't happen with stress test failures. Exact same behavior every time, no BSODs. Been overclocking for a long time and never seen this before.

 

[Idontcare]

You may try cleaning your HSF, dust can dramatically reduce the cooling efficiency of the heatsink. Another thing to try is replacing the TIM as TIM's can dry out or get moved around (physically) out from under the HSF owing to the so-called "pump-out" effect.

I had this problem on my laptop, CPU fan was going to 100% all the time. I removed the laptop cooler and replaced the TIM and now the CPU doesn't even come on loud enough for me to hear.

 

[evilqueens]

Hello again,

So I finally got around to installing my NH-D14 and I noticed that my load temps (for stock settings - 3.9GHz turbo) are about 10-15ºC lower than with the stock Intel CPU heatsink.

I pushed my overclock to 4.5GHz with a 1.25V VCore and LLC set to Ultra high (75%). With this setting, I managed to get a stable overclock with LynX and my temperatures were spiking at 85ºC for the hottest core.

I'm just wondering if this is reasonable? Is it okay to have my chip running at this setting 24/7? Theoretically, I don't think there's anything I'd ever do that would keep my temperatures this high for normal usage. But then again, I'd rather not have it at a setting that's going to cause damage over time.

Anyway, just wanted your opinions - thanks!

My system:

Intel Core i7-3770K CPU
ASUS Maximus V Formula Motherboard
G.Skill Ripjaws X 16GB (4x4GB) DDR3-1600 CL9 RAM (9-9-9-24 CR2) @ 1.5V
Gigabyte GeForce GTX 680 OC GPU
Corsair HX850 Power Supply

 

[Idontcare]

Hello again,

So I finally got around to installing my NH-D14 and I noticed that my load temps (for stock settings - 3.9GHz turbo) are about 10-15ºC lower than with the stock Intel CPU heatsink.

I pushed my overclock to 4.5GHz with a 1.25V VCore and LLC set to Ultra high (75%). With this setting, I managed to get a stable overclock with LynX and my temperatures were spiking at 85ºC for the hottest core.

I'm just wondering if this is reasonable? Is it okay to have my chip running at this setting 24/7? Theoretically, I don't think there's anything I'd ever do that would keep my temperatures this high for normal usage. But then again, I'd rather not have it at a setting that's going to cause damage over time.

Anyway, just wanted your opinions - thanks!

My system:

Intel Core i7-3770K CPU
ASUS Maximus V Formula Motherboard
G.Skill Ripjaws X 16GB (4x4GB) DDR3-1600 CL9 RAM (9-9-9-24 CR2) @ 1.5V
Gigabyte GeForce GTX 680 OC GPU
Corsair HX850 Power Supply

Those numbers look fine, I see nothing to be concerned about.

This is without delidding, right? If yes, then your numbers are great.

 

[evilqueens]

Those numbers look fine, I see nothing to be concerned about.

This is without delidding, right? If yes, then your numbers are great.

Yeah - I haven't de-lidded it yet. Based on your numbers, I may do it soon, though!

Regarding fine-tuning the overclock, does it make more sense to increase LLC for a given voltage, or to bump the voltage with no LLC?

For example, I was getting BSODs at 1.225V VCore with LLC set to medium (25%). I finally increased it to 75% to get a stable overclock - but this means my actual VCore under load is actually about 1.233V. (Sorry, I made a mistake in my other post - my VCore is manually set to 1.225 not 1.25)

So, does it make sense to set the LLC to ultra high (with VCore at 1.225V), or to just manually set the VCore to 1.235V and be done with it.

On another note, I still can't get a 7.9 on my WEI page. Ack!

 

[Idontcare]

Yeah - I haven't de-lidded it yet. Based on your numbers, I may do it soon, though!

Regarding fine-tuning the overclock, does it make more sense to increase LLC for a given voltage, or to bump the voltage with no LLC?

For example, I was getting BSODs at 1.225V VCore with LLC set to medium (25%). I finally increased it to 75% to get a stable overclock - but this means my actual VCore under load is actually about 1.233V. (Sorry, I made a mistake in my other post - my VCore is manually set to 1.225 not 1.25)

So, does it make sense to set the LLC to ultra high (with VCore at 1.225V), or to just manually set the VCore to 1.235V and be done with it.

On another note, I still can't get a 7.9 on my WEI page. Ack!

In regards to LLC - there are two answers: (1) the pendantic, academically correct to the fifteenth decimal, answer, and (2) the practical, good-enough-for-government-work, answer.

I use to be worried about (1), the academically correct answer that was fit for college textbooks. And 4-5 yrs ago when LLC was in its infancy and the raw implementations were rather crude, it made sense to worry about it.

But now, with the technique becoming so mature and functional, it really isn't worth the physical time or the mental effort IMO to worry about the finer points of micro-optimizing LLC settings.

As far as I can tell, at the practical level it is all six of one or a half-dozen of the other.

Whether you give the cores the volts they need for stable operation by way of tweaking the LLC settings or just by way of manually brute-forcing the Vcc value such that the cores are still stable despite the Vdroop, both are effective approaches and the practical differences in the CPU's expected lifespan are irrelevant for anyone who plans to retire the CPU in less than 10yrs anyways.

For me, just because it seems easiest to do in practice, I set my LLC to 50% or 75% (depending on the situation) and dial in the Vcc value needed for stability and call it day.

 

[Idontcare]

Yeah - I haven't de-lidded it yet. Based on your numbers, I may do it soon, though!

I generated some numbers for my 3770k and put them in a graph to highlight the striking advantages that can come from reducing your CPU's operating temperature as much as possible:

^ By reducing the operating temperature you reduce the power usage (reduction in static leakage) from the horizontal purple line the declining green line. For my chip at 4.6GHz this reduction in power was around 12% if I did not reduce the voltage at all, just reduced the temperature.

Furthermore, by reducing the operating temperature you also reduce the voltage that is required by the CPU for stable operation at any given clockspeed.

This reduction in the required voltage then goes towards further reducing the power consumption (reduces both leakage and dynamic power), dropping the CPU power consumption from the green line to the red line in the graph above.

^ all told, by reducing the temperature I was able to reduce the voltage needed for stable operation by some 5%, netting a 21% reduction in power consumption by the CPU all because I reduced the operating temperature by some 40°C through a combination of replacing the stock HSF with a lapped H100 and by delidding the IB chip :thumbsup:

 

[evilqueens]

Oooh! You are making it all the more tempting to do it!

I think I'll hold off for now - I see no discernible need to run my CPU past its stock frequencies at the moment.

I'm quite happy with 3.9GHz @ 0.984V Vcc (under load with a peak core temperature of 59ºC after 5 cycles of LinX).

When it starts to get old, or I need to push it, I'll probably de-lid it then.

On another note, do you know if its normal to be able to touch the Noctua heatsink directly without it feeling hot? I was able to put my fingers directly on the aluminum heat pipes near the plate that makes contact with the CPU and it didn't feel hot to touch at all (just a little warm).

This is strange, because 60ºC is a temperature that can give you pretty serious burns!

 

[Idontcare]

Yeah that is normal.

The temperature down in the core of the cpu is much lower than the temperature way up near the surface of the HSF mount.

But you can get the fins hot by taking the fans off and letting the CPU heat all the way up to TJmax. The air around the fins will be stagnant and hot, and the fins will get hot too.

^ the red dot you see on the side of the NH-D14 is a laser beam that shows where the IR temperature reading is being taken at. In this test the fins got to 63.5°C.

When I put the fans back on, the CPU temp drops rapidly from TJmax and the NH-D14 fins fall to nearly ambient temperatures (30-35C).

 

[evilqueens]

Ah I see. Interesting that such a high temperature differential (fin temperature to core temperature) can exist when essentially it's all metal (and some TIM) connecting everything.

How can the Noctua heatsink plate surface be cool enough to touch, when the core (and CPU package) temperatures get as hot as 60-80ºC?

Let's hope that Intel can come up with some better temperature characteristics as they scale down their process size even further in the future.

Anyway, I just re-ran some tests and this is what I got:

3.9 GHz @ 0.984 Vcc measured (Offset mode with - 0.175V offset) - maximum peak core temperature: 59ºC
4.4 GHz @ 1.175 Vcc measured (Offset mode with - 0.050V offset) - maximum peak core temperature: 74ºC
4.6 GHz @ 1.295 Vcc measured (1.185V fixed Vcc in BIOS + 75% LLC) - maximum peak core temperature: 89ºC

Even though I could, I don't think I want to push it any higher than that for now. I'm worried about permanent damage, and I was wincing when I let LinX run at 4.6GHz and saw the temperatures almost hit 90ºC

Anyway, it was a great experiencing learning how to overclock and all. Definitely glad I took some time to do it. But for now, it's back down to 3.9GHz for me!

 

[Idontcare]

How can the Noctua heatsink plate surface be cool enough to touch, when the core (and CPU package) temperatures get as hot as 60-80ºC?

It is all about the concentration of the heat - temperature is heat density.

Inside the cores, the space is tiny but all that heat is contained in there. So the local temperature is high.

Outside the cpu, into the IHS and the HSF block, the volume of metal is quite large, so the same amount heat divided into a larger volume results in a lower temperature.

Think about when you cook food on your stove. That burner is red-hot (or blue hot if you use gas)...but the pot sitting on the burner doesn't become red hot or blue hot because the pot is at lower temperature despite being in direct contact with the burner or flame.

 

[evilqueens]

Yes, this is true. I did forget that the die size is actually quite a bit smaller than the IHS. Then throw on the huge heatsink (and the two fans running full-blast), and it definitely makes sense that it is cooler.

 

[Ed1]

Ah I see. Interesting that such a high temperature differential (fin temperature to core temperature) can exist when essentially it's all metal (and some TIM) connecting everything.

How can the Noctua heatsink plate surface be cool enough to touch, when the core (and CPU package) temperatures get as hot as 60-80ºC?

Let's hope that Intel can come up with some better temperature characteristics as they scale down their process size even further in the future.

Anyway, I just re-ran some tests and this is what I got:

3.9 GHz @ 0.984 Vcc measured (Offset mode with - 0.175V offset) - maximum peak core temperature: 59ºC
4.4 GHz @ 1.175 Vcc measured (Offset mode with - 0.050V offset) - maximum peak core temperature: 74ºC
4.6 GHz @ 1.295 Vcc measured (1.185V fixed Vcc in BIOS + 75% LLC) - maximum peak core temperature: 89ºC

Even though I could, I don't think I want to push it any higher than that for now. I'm worried about permanent damage, and I was wincing when I let LinX run at 4.6GHz and saw the temperatures almost hit 90ºC

Anyway, it was a great experiencing learning how to overclock and all. Definitely glad I took some time to do it. But for now, it's back down to 3.9GHz for me!

Running with 0.175v offset , how or low does it go when your running 1600mhz at idle, or you have speedstep turned off ?

Yes, HS don't get hot at all , my 212evo under load only goes to 85F anywhere measured on HS .Those core temps are internal temps so way higher than what the IHS/die is at surface by HS base .

 

[evilqueens]

Running with 0.175v offset , how or low does it go when your running 1600mhz at idle, or you have speedstep turned off ?

Yes, HS don't get hot at all , my 212evo under load only goes to 85F anywhere measured on HS .Those core temps are internal temps so way higher than what the IHS/die is at surface by HS base .

I think the lowest I've gotten it at 1.6GHz is 0.705 V VCore. I actually really like speed step - I don't really know why you'd turn it off. Maybe if you need some crazy fast response times - but I've never run into a situation where the normal throttling up/down hasn't been fast enough.

 

[Joey Bogaars]

Hi everybody,

I bought a Compaq CQ57 with a E-450+Radeon 6320 and to be onest, this system is unstable and rubbish, i've tried everything to make it better but it doesn't work at all. The CPU get stuck when im multitasking and the screen is hanging when i play youtube-clips/video's. This is kinda annoying, the CPU is way TOO instable, the Radeon 6320 is quite good...the graphics seems to be good enough if it doesn't lag. But the system overall is, not good at all. I have a tweaked Windows 7 x64 Ultimate and even that doesn't work. I replaced my HDD for a SSD and even that doesnt work, i bought 2GB extra RAM but even that seems not to bring every difference. AMD-Overdrive doesn't work on this CPU/GPU. What to do? Overclocking is impossible, more RAM isn't a option. Please help me out and give me some options.

I've expected more of this E2 system of AMD.

Gr. Joey

 

[dini92]

 

[Markfw]

Hi everybody,

I bought a Compaq CQ57 with a E-450+Radeon 6320 and to be onest, this system is unstable and rubbish, i've tried everything to make it better but it doesn't work at all. The CPU get stuck when im multitasking and the screen is hanging when i play youtube-clips/video's. This is kinda annoying, the CPU is way TOO instable, the Radeon 6320 is quite good...the graphics seems to be good enough if it doesn't lag. But the system overall is, not good at all. I have a tweaked Windows 7 x64 Ultimate and even that doesn't work. I replaced my HDD for a SSD and even that doesnt work, i bought 2GB extra RAM but even that seems not to bring every difference. AMD-Overdrive doesn't work on this CPU/GPU. What to do? Overclocking is impossible, more RAM isn't a option. Please help me out and give me some options.

I've expected more of this E2 system of AMD.

Gr. Joey

spyware or viruses ? sounds like it to me.

 

[lehtv]

Just posting my findings regarding LinX vs Prime95 stability testing because it contradicts Idontcare's OP which claims LinX is better for determining core logic stability. I'm wondering if I'm doing something wrong or if one program is better for lower voltages while the other is better for higher voltages.

I've been tinkering around with offset voltage, trying to find the minimum 24/7 stable voltage for 4200MHz (I want to keep it quiet and cool at full load). In short, LinX allows me to use a much lower voltage than Prime95 Small FFT. Here's what I did:

Initial settings:
- offset voltage -0.005
- turbo multi 42
- LLC 0% (regular)
- HT off
- all power saving functions normal

In LinX, I set memory to [All]. After each completed cycle, I decreased voltage by 0.005 in Asus AI Suite II until it would crash. Between each cycle, the thread loads and temperatures decreased substantially for a short while - is this normal? I got as low as -0.070 offset. Upon booting, I increased the offset to -0.055 and left it testing a four hour period which should amount to around 80 cycles or so. I got no errors during that time.

Now, if I enable HT and test Prime95 Small FFT's with the same voltage that I just passed 4 hours of LinX with, it BSOD's within a few seconds. I'm now running -0.040 but I've yet to determine exactly what voltage I need for overnight stability in Small FFT's and large FFT's, gonna leave it churning this evening.

See anything wrong with that? Regarding my choice to set HT off instead of assigning thread affinities in Task Manager... well, I couldn't get LinX to properly stress the four threads I chose. I tried threads 0/2/4/6 as well as 1/3/5/7, but thread 6 or 7 would always stay closer to 0% than 100% load during stability testing.