GK104 and voltage

Can somebody explain to me when or if it's ok to adjust voltage on a GK104? I only have a 660ti, but it seems to clock decently. Power target 123%, GPU Offset +145, and Mem Offset +500 were stable in Kombuster/P95 (same time), but I got a few crashes in Crysis 2 and started crunching on GPUGrid recently and was getting failed results.

I set the card back to default (it's a 100% factory reference card, PCB and clocks) to see if that helps the crunching errors, but I'll have to wait a day or so to see if stock clocks return valid results.

I know GK104 has thermal and power constraints, but at what point is ok to add a little voltage? I've heard that adding voltage might make the card not boost as much, but I don't know what the truth of the matter is. I had trouble finding a really good guide for Kepler.

GPU temps are fine. Card has a MWC82 water block, heatsinks on the RAM, and the stock VRM heatsink with a low speed fan blowing over the whole thing. Default voltage in PrecisionX is 987 mV, should I bump it a bit (how much?) or leave it be and scale back the clock offset a little?

Unless you're using some kind of modified BIOS that allows for higher GPU core voltages than NVIDIA allows, I would just raise the voltage as high as it allows. Then starting at default clock speed, begin bumping it up in increments. One of the best ways I've found to test stability is by running HWBot Heaven benchmark, it really pounds on a GPU. Once you find the highest stable GPU core speed, then start working on the memory. If you still get gaming and crunching errors after you do this, then start by lowering the memory a bit. If you still get errors, then the GPU core speed needs to be dropped.

Lvcoyote said:

HWBot Heaven benchmark

Click to expand...

I've used Heaven in the past, but Kombuster seemed to show problems faster. Is the HWBot version different somehow, or is it just Heaven 4.0 with some specific preset?

Probably the same. All you can really do then is play around with the core and memory speeds until you find the fastest stable clocks.

Alright, so max the power and voltage sliders and then work from there. Sounds like a plan.

Coming from a reference GTX570, I never wanted to touch the voltage slider, lol. I knew Kepler handled it very differently, but I didn't know you could just max the voltage like that without worry.

You can also open up GPU-Z and using the sensor function you can save that information to a txt file for viewing. You can look at that text file and see if the GPU is throttling down while the stress test is running.

Yea, I'm noticing that Kombuster just causes the card to downclock to 915 MHz if I max the voltage slider. It does get the GPU good and toasty, but it's not a good stability test if the card won't clock. TDP hits 132% according to GPU-Z, which is curious, but I guess that's why the card is clocking to 915 MHz despite offsets. Clocks normally with the same settings in Heaven, but doesn't get as hot or consume as much power.

This is going to require some tinkering. Luckily I kind of like tinkering. Thanks for the info.

All GTX600 series have max safe voltage of 1.25V but depends from manufacturer and used voltage controllers it may vary. All what you see in BIOS/software are VIDs , not real voltage so best is to check it under load with multimeter.
Regardless VIDs etc, max in BIOS/software ( except maybe series like Lightning etc ) is 1.212V so it won't make more than these 1.25V. OCP/OVP is usually set to ~1.28V+ so it shouldn't cause throttling or black screens etc.

I would get Kepler BIOS Tweaker 1.25 and edit BIOS. Disable boost, unlock max voltage to 1.21V, set higher max 3D clock or raise power limit.
I doubt it's the power limit and for me it looks more like a regular nvidia boost issues. I had about the same on GTX670 but I was running at ~1200 core clock.
GPU-Z is often showing wrong values so depends from card and if you are using modded BIOS, read can be good or wrong. Like when you change power limit on already modded BIOS then GPU-Z won't show you 130%+ as base but 100%. The same with voltages. When you raise voltage then GPU-Z will still show you 1.15V even though card can have 1.2V+ ( because it reads VID, not actual voltage ).

Downclocking can be also caused by drivers. Older drivers were causing throttling at 80+*C. Some new series are starting at as low as ~60 or 70*C. Try older drivers like ~310 series and check if clock is acting the same.
In Kepler BIOS Tweaker you can set minimum and maximum clock to the same value then it won't throttle at all but it will use more power.