Advanced search

Message boards : Number crunching : Overclocking: what do I tweak first?

Author Message
Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28223 - Posted: 26 Jan 2013 | 20:11:01 UTC

I have my GTX 570 running on manual fan control and running between 40C and 50C so I want to try a little OC. I've read advice in these forums about what to tweak first and by how much but I can't find it or remember. Right now my clocks are:

Processor: 1484 MHz
Graphics: 742 MHz
Memory: 1900 MHz

IIRC one should boost the processor clock first? What is a reasonable boost in my case... 1%, 2%, 10%? And if tasks then start to fail one should tweak the memory clock up a wee bit too? Or am I getting this confused with OCing CPUs (which I have never done either, btw, just read about it)

I've also read here that it's difficult to regain time wasted by OCing too much and getting compute errors and results that won't verify so I intend to take this slowly and cautiously.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28224 - Posted: 26 Jan 2013 | 21:26:33 UTC - in response to Message 28223.

I have my GTX 570 running on manual fan control and running between 40C and 50C so I want to try a little OC. I've read advice in these forums about what to tweak first and by how much but I can't find it or remember.

In this case use the "advanced search" function with no time limit.

Right now my clocks are:

Processor: 1484 MHz
Graphics: 742 MHz
Memory: 1900 MHz

IIRC one should boost the processor clock first?

Yes. Note that the Graphics and the Processor (aka shader) clocks are not independent, their ratio (x2) is fixed.

What is a reasonable boost in my case... 1%, 2%, 10%?

10%-15% is a reasonable boost using the factory made cooling.

And if tasks then start to fail one should tweak the memory clock up a wee bit too?

No, the GPUGrid tasks are not GPU memory intensive, no gain in speed or stability could be achieved by increasing the graphics memory clock.
You should increase the GPU voltage by 25mV when tasks start to fail, as long as your GPU temps are below 80°C (and you can tolerate the noise of the cooling). In the other case you should decrease the GPU clock by 10MHz.

Or am I getting this confused with OCing CPUs (which I have never done either, btw, just read about it)

Yes. Actually the increase of the memory clock is needed because the CPUs have so high CPU and memory multipliers, and it's practical to keep them as low as possible. (and the standard CPUs have a limit on these multipliers)

I've also read here that it's difficult to regain time wasted by OCing too much and getting compute errors and results that won't verify so I intend to take this slowly and cautiously.

That's the good approach.
First you could increase your GPU clock rate by 10%, and the GPU voltage by 50mV. If it's stable, you can increase your GPU clock by 10MHz increments as long as it stays stable (crunches without errors for a week). Then you should follow the stability regain procedure I've described above. Note that different workunit batches tolerate different overclocking, so if your overclock settings have no "safety overhead" (ie. a little higher GPU voltage than needed, or a little lower GPU clock than the highest your GPU can run at the given voltage), then you should recalibrate your overclocking settings when tasks start to fail.

Dylan
Send message
Joined: 16 Jul 12
Posts: 98
Credit: 386,043,752
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwat
Message 28229 - Posted: 27 Jan 2013 | 0:51:12 UTC

How possible do you think overclocking is on a GTX 670, I've heard Nvidia locked the voltage. Is it better to simply run stock to avoid WU failures?

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28230 - Posted: 27 Jan 2013 | 1:13:45 UTC - in response to Message 28224.
Last modified: 27 Jan 2013 | 1:15:50 UTC

OK, then the plan is as follows:

First you could increase your GPU clock rate by 10%, and the GPU voltage by 50mV. If it's stable, you can increase your GPU clock by 10MHz increments as long as it stays stable (crunches without errors for a week). Then you should follow the stability regain procedure I've described above. Note that different workunit batches tolerate different overclocking, so if your overclock settings have no "safety overhead" (ie. a little higher GPU voltage than needed, or a little lower GPU clock than the highest your GPU can run at the given voltage), then you should recalibrate your overclocking settings when tasks start to fail.


The only trouble is I cannot see how to adjust voltage. I am using Linux. I can see how one uses the nvidia-settings command plus appropriate options/arguments to query and adjust clocks and fan speed but I see nothing for voltage. I did the following 2 commands:

nvidia-settings -q all | grep vcore
nvidia-settings -1 all | grep volt


I did
nvidia-settings -q all > queries.txt


and used the text editor to search for anything to do with volts and found nothing so... how do I adjust voltage.

In this post Carlesa25 seems to say you need Coolbits 1 to adjust the voltage and Coolbits 4 to adjust the fan speed and that you cannot have Coolbits 1 and Coolbits 4 simultaneously. So I changed Coolbits 4 to Coolbits 1 and rebooted but still there is nothing in nvidia-settings -q all about voltage.

How can I tweak the voltage?

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28241 - Posted: 27 Jan 2013 | 12:51:33 UTC - in response to Message 28229.
Last modified: 27 Jan 2013 | 13:12:39 UTC

How possible do you think overclocking is on a GTX 670, I've heard Nvidia locked the voltage. Is it better to simply run stock to avoid WU failures?

I'm a little confused about overclocking the GTX 670, because this series has a 'GPU power' measurement, and it's not clear to me what this measurement is about?
I have two Asus GTX670-DC2OG-2GD5s, and a GigaByte GV-N670OC-2G. Both types are factory overclocked a little, the GigaByte has larger PCB, mostly because it has a more powerful PSU on board.
According to the above, their 'GPU power' measurements are very different:

1. Asus GTX670-DC2OG-2GD5: GPU power: ____ 97%, core voltage: 1.175V, core clock: 1084MHz
2. Asus GTX670-DC2OG-2GD5: GPU power: 104-107%, core voltage: 1.162V, core clock: 1110MHz
3. GigaByte GV-N670OC-2G:_ GPU power: ____ 72%, core voltage: 1.175V, core clock: 1150MHz

As you can see, according to the GigaByte's GPU power reading it's consuming much less than the Asus', while the chip used on these cards couldn't be different that much. So this reading must be the ratio of the available power from the PSU on the card and the power the chip is actually using. If this is correct, then a reading like 107% scares me :) and I don't want to overclock those cards. However, I've overclocked the GigaByte:

4. GigaByte GV-N670OC-2G:_ GPU power: ____ 75%, core voltage: 1.175V, core clock: 1240MHz

While I've set the MSI Afterburner to raise the GPU voltage by 50mV, it didn't happened. Also, I've set +100MHz, and the real increase is only 90MHz, so I'm still a little confused.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28246 - Posted: 27 Jan 2013 | 14:34:05 UTC - in response to Message 28241.

I thought it was a percentage of the TDP, however I wasn't sure if it was a percentage of the reference value or the bespoke GPU's value (which I don't know anyway)! I can't imaging a FOC'ed GPU would operate 'up to' the reference TDP.

I haven't been able to OC my GTX660Ti using MSI Afterburner. I suspect this is because I had another GPU in the same system, and swapped them into different slots a few times.

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28247 - Posted: 27 Jan 2013 | 14:41:19 UTC
Last modified: 27 Jan 2013 | 14:42:23 UTC

I Googled "Linux overclock nvidia" and found many articles stating Fermi cards cannot be overclocked on Linux. Unless things have changed with the 313.x driver I think I must conclude the only way to overclock nVidia with Linux is to hack the card's BIOS similar to the way it's described here. Note they did the hack on a GT220. What's the odds of it working on a GTX570? Has anybody done this successfully?

Dylan
Send message
Joined: 16 Jul 12
Posts: 98
Credit: 386,043,752
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwat
Message 28248 - Posted: 27 Jan 2013 | 19:04:48 UTC

I found a really great OC guide that I was reading. I don't know how much it will help, but I will post it anyway. It does touch on the power measurement, and includes, at the end of it, links to OC guides for specific vendor cards like the Asus 670 DC2T and what not.


http://www.overclock.net/t/1265110/the-gtx-670-overclocking-master-guide

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28267 - Posted: 28 Jan 2013 | 19:47:48 UTC - in response to Message 28248.

I found a really great OC guide that I was reading. I don't know how much it will help, but I will post it anyway. It does touch on the power measurement, and includes, at the end of it, links to OC guides for specific vendor cards like the Asus 670 DC2T and what not.


http://www.overclock.net/t/1265110/the-gtx-670-overclocking-master-guide

Nice reading! It will take a couple of days to read, understand and try it all.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28271 - Posted: 28 Jan 2013 | 22:31:51 UTC - in response to Message 28267.

Somewhat OT but I was not able to OC, possibly because I had more than one generation of GPU in the same system, had swapped them around at least 3 times and upgraded drivers at least once (W7). Removed one card but when I tried to move from 306.97 to 310.90 I ended up at a black screen, the monitor's power LCD was red - it was off. The computer was on, but I couldn't use the monitor.
I had to start up in safe mode, uninstall as much of the NVidia drivers as I could, reboot, finish uninstalling the drivers, and MSI Afterburner, restart, install the latest driver, reboot and then install MSI Afterburner.
I was then able to OC the GTX660Ti to 1250MHz ( ͡° ͜ʖ ͡°)
I'm not running it here just yet, as the power usage is higher here and I'm not keen on it crashing overnight during a long run. The core clock (min and max) made no sense in MSI afterburner, so I increased it (and the other settings) slowly and looked at GPUZ to see what I was actually doing. My GPU says 1111MHz is the Boost rate (a 5% FOC), so 1250MHz isn't too much more (12.5% over the FOC) - it might work here, as long as I keep it below 70°C (which shouldn't be an issue with the GTX470 pulled, for now). I do like the idea of a GTX660Ti matching a stock GTX680...
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Dylan
Send message
Joined: 16 Jul 12
Posts: 98
Credit: 386,043,752
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwat
Message 28274 - Posted: 29 Jan 2013 | 0:10:00 UTC - in response to Message 28271.
Last modified: 29 Jan 2013 | 0:17:11 UTC

I was then able to OC the GTX660Ti to 1250MHz
My GPU says 1111MHz is the Boost rate (a 5% FOC), so 1250MHz isn't too much more (12.5% over the FOC) - it might work here, as long as I keep it below 70°C (which shouldn't be an issue with the GTX470 pulled, for now). I do like the idea of a GTX660Ti matching a stock GTX680



Get back to this thread on the results of your overclocking. I overclocked my two EVGA 670's to about the same clock speeds and a few tasks were failed recently, so I decided to put them back to stock. You also said you used Afterburner? I use EVGA's Precision program, do you think that would make a difference?

Also, one of my cards is running at about 71-72 C, even though the fan speed is as far is it can go (80%) and I placed an addition small fan (not a case fan, a fan about the size of a table lamp) on the side of the case to blow air directly onto the GPUs, and the ambient room temperature is about 26 C. I was wondering how you keep your cards under 70 degrees. Perhaps I should start a separate thread about this.

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28275 - Posted: 29 Jan 2013 | 1:55:28 UTC - in response to Message 28274.
Last modified: 29 Jan 2013 | 2:02:55 UTC

If starting a thread confers ownership of it then I could say I own the thread but I don't think that way. As far as I am concerned you're on topic here, FWIW. Heat issues and overclocking are all part of the same topic in my mind.

I've done a fair bot of work on the heat and temperature thing, I think. On Linux I cannot get the fan to go any higher than 85% and several other sources indicate that is the max allowed by nVidia in the BIOS. Windows may be different but you may be able to squeeze another 5%. I've noticed if I set the fan to 100% it will fall back to ~45%. If I set it to 85% it will fall back to ~50% which is very weird but the point is if you set it to just the right number it will fall back to 85%, at least that's the way it works on Linux.

I doubt an additional 5% will keep your card under 70C if ambient room temp is 26. I don't have any graphs or tables to base this estimate on but from a little tinkering and observing I would say you could get under 70C with an ambient room temp of ~20C. To achieve that 20C you need to consider venting your GPU to the outdoors or into the house attic or any room but the computer room. The alternative would probably be air conditioning.

I have a highly customized cooling solution and I am working on clearing some other tasks from my plate so I can put together a video or series of pics to show how easy it is to engineer and create your own very capable cooling solutions. What you need to do is adapt a piece of duct to the back of your computer case where the GPU vents in order to catch that hot air and carry it away. You can do that easily with a pair of tin snips, a measuring tape and a little knowledge of simple geometry. Actually if you buy common aluminum duct you can cut that with a decent pair of scissors, expensive tin snips not required. To fasten your creation in place you can literally use duct tape and string and glue, or you can drill holes and fasten with pop rivets or screws. You canget a decent drill for about $20 brand new, better deals at the pawn shop if you talk them down and they always go down. If you buy self tapping screws you don't even need to buy drill bits though you'll want them eventually if you have a drill.

Once you have that adaptor the rest is easy... just connect elbows and straight pieces to carry the air out the window or through a hole in the ceiling into the attic, through the floor into the basement, wherever you wish. And you can put inexpensive helper fans into the duct to assist the flow over longer runs though with standard and inexpensive 4 inch clothes dryer duct you need not worry too much about that.

Yeah it takes tools but you can invest in tools and make stuff for yourself cheap or you can blow money on expensive off the shelf solutions that often don't work half as well as something you design to meet your specific need. Whatever creams yer twinky.

Also, I'm considering either hacking the BIOS on the card to allow fan speeds higher than 85% or else rewiring the fan to force it to run at 100% which has obvious benefits and may negate the need for ducts or lower ambient temp. That requires a soldering iron and confidence as you would need to snip wires and resolder stuff on the card. Steady hands helps too :-)

Hmmm, steady hands and confidence... isn't that why the gods gave us whisky?

werdwerdus
Send message
Joined: 15 Apr 10
Posts: 123
Credit: 1,004,473,861
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28276 - Posted: 29 Jan 2013 | 3:00:50 UTC

on all of my 660 Ti cards I used MSI Afterburner and maxxed out the Power Limit slider (113%) and it automatically increased the boost gpu clock. If it is possible to increase the Power Limit I would try that first, as it would (AFAIK) increase the voltage a little bit if needed to reach a higher boost clock.
____________
XtremeSystems.org - #1 Team in GPUGrid

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2343
Credit: 16,201,255,749
RAC: 0
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28292 - Posted: 29 Jan 2013 | 23:37:22 UTC - in response to Message 28276.
Last modified: 29 Jan 2013 | 23:38:32 UTC

Oh, I've find out, that NVidia forbid the card manufacturers to enable voltage tweaking of the Kepler chips above 1.175V. The MSI Afterburner V2.23 is the last version which supports voltages above this limit. Unfortunately it doesn't support GTX660(Ti).

Dagorath
Send message
Joined: 16 Mar 11
Posts: 509
Credit: 179,005,236
RAC: 0
Level
Ile
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28294 - Posted: 30 Jan 2013 | 0:31:31 UTC - in response to Message 28292.

Is 1.175 volts enough for a decent OC? If not then what is the alternative, is it necessary then to modify the BIOS to get above 1.175?

Dylan
Send message
Joined: 16 Jul 12
Posts: 98
Credit: 386,043,752
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwat
Message 28295 - Posted: 30 Jan 2013 | 0:51:30 UTC

It should be ok. Search for what other people have achieved with your GPU and see if that is ok for you.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28307 - Posted: 30 Jan 2013 | 14:40:00 UTC - in response to Message 28295.

I tried 1250 and 1220MHz for the GTX660Ti, but both OC's resulted in Driver error pop-up messages (310.90). I suspended the tasks, closed the pop-up and was able to resume at default clocks. The TDP was around 85%, the fans were reasonably low and the temps were low ~60°C. The card can run other projects at 1250MHz but not GPUGrid tasks. I guess the 1.175V cap is causing this problem. Other projects don't use as much power.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28308 - Posted: 30 Jan 2013 | 15:03:01 UTC - in response to Message 28307.

I've slightly changed my approach to this, and might be getting somewhere:
Firstly, it could well have been the case that using GPUz that was causing my driver failure - seen that before! So I'm steering clear of GPUZ.

I also decided not to use a fan profile and fixed the GPU fans at 75%. This resulted in a drop of temperature to 55°C.

I then increased the Power limit to 107%

Knowing that the GPU only goes up in steps of 13MHz I bumped it up by 26MHz and the GPU ran at 1215MHz, and used 89% of the TDP. I know this isn't much over what I was usually getting 1175, 1188 or 1201MHz, but it's stabilized and a bit of an OC.

I then increased the Power limit to 109% and the GPU to 1228MHz. The Power usage went up to ~90% of the TDP. So far 1228MHz looks stable.

____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

werdwerdus
Send message
Joined: 15 Apr 10
Posts: 123
Credit: 1,004,473,861
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28312 - Posted: 30 Jan 2013 | 18:14:12 UTC

remember that the 660 Ti will "boost" its clock when under high GPU usage. So if you are setting a certain MHz, it will actually go higher than that when under load. The higher power limit will allow that boost to go higher than normal. Something to keep in mind when setting MHz on the core/shader. You may have to decrease the power limit if you are raising the core clock, or lower the core clock if you are raising the power limit.
____________
XtremeSystems.org - #1 Team in GPUGrid

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 28319 - Posted: 31 Jan 2013 | 18:35:32 UTC - in response to Message 28312.

Ended up with a failure, after 7 or 8h, but I suspect that these NOELIA tasks are troublesome, the same task failed on two other systems (albeit early on), and plenty of others are also seeing Energies have become nan errors too. I don't believe a mildly overclocked GPU is to blame, especially running at 55°C.

The same card has since completed another task.





____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Post to thread

Message boards : Number crunching : Overclocking: what do I tweak first?

//