~~The GTX 670 Overclocking Master-Guide~~

The GTX 670 Overclocking Master-Guide

Introduction:
With the mass-popularity of the new Nvidia GTX 670, and the constant influx of questions related to overclocking them, I've decided to write an all-inclusive master-guide to overclocking them that should help most people get off to a strong start with their new 670's. All of the Kepler-based GPUs (670, 680, and 690's) are a very unique breed of GPU. Gone are the days of manually increasing voltage to stabilize an otherwise unstable overclock. Now, the user must use a great deal of finesse, and a ton of trial and error, to maximize the potential of their overclock. We now have to worry about dynamic clocking, dynamic volt changes, temperatures, and power draw in-order to reach a maximum stable overclock.

Table of Contents:

Vocabulary:

• GPU Clock
• Boost Clock
• Kepler Boost
• Max Boost Frequency
• Power Percent/Power Target
• GPU/Memory Clock Offset

Software: Benchmarks and Overclocking Tools:

• Drivers
• Precision-X
• GPU-Z
• Unigine 'Heaven' DX11 Benchmark
• 3DMark 11 Benchmark
• Furmark
• Other Free Game Benchmarks

Overclocking:

• The Long Method
• The Quick Method
• Setting up a Quiet and/or Efficient Fan Curve

Troubleshooting:

• How to Receive Help With Your Overclock in This Thread
• Reducing Temperatures to Eliminate Throttle
• FAQ

Other:

• Videos
• Additional threads
• Closing Notes

The guide has now been translated to Italian over at hwupgrade by johnnystuff

---

Change-log: Warning: Spoiler! (Click to show)
>>Link back to the table of contents<<
Edited by SeanPoe - 6/27/12 at 7:36pm

Vocabulary:
Before we begin, I feel it's important to go over some new vocabulary that's unique to the Kepler-based GPUs. With all of the new architecture and functionality changes brought about with the GTX 670, I think it will be very beneficial for the community as a whole if we, the GTX 670 owners, could come up with some common terminology. This is a step towards that goal.

GPU Clock:
In the past, the GPU clock was the maximum and minimum frequency the core of the GPU would run at under full-load. It had significance, and it was the number people would look at when judging the performance of a GPU. Now, with the Kepler-based GPU's dynamic clocking, this number has little to no real-world value. For all intents and purposes, this number can just be ignored.

Boost Clock (abbreviated as "BC"):
This number is quite similar to the original GPU clock of past generations, with some key differences though. The first key difference is the Boost Clock is not the maximum core speed that the GPU will run at under load. However, it is still typically the minimum core speed that the GPU will run at under full-load. The second key difference is this number alone can not be used to judge the general performance of a GPU.

How to find your Boost Clock: Warning: Spoiler! (Click to show)
Kepler Boost (aka, Dynamic Clocking Boost, abbreviated as "KB"):
This is the new Dynamic Clocking feature found on the Kepler-based GPUs; It will dynamically add more performance to the card if it's at a suitable temperature and suitable power percent. If the GPU is not drawing too much power or running too hot, then this Kepler boost value is added to the Boost Clock when the card is under load. It's maximum amount is set in stone and can't be increased beyond what Nvidia set it at when the GK104 chip was made. Each card has a different maximum Kepler Boost. For example, some cards will have a KB value as low as 104MHz (and perhaps even lower) and some as high as 234MHz, the average is somewhere around 130MHz and 143MHz. The only constant is the Kepler Boost is always a multiple of 13 (ie, 13, 26, 104, 130, 234, etc). However, the thing to keep in mind is, the maximum Kepler Boost will only be active when the card is not too hot or drawing too much power. If the card gets too hot, then the Kepler Boost value will be throttled down below its true maximum in 13MHz increments. The first thermal throttle point is at 70C, then 80C, then 85C and lastly at 95C. Here's a graph for easy reference:


Additionally, If the card is drawing too much power, it will also throttle itself down in 13MHz increments. I will go into more detail below on how to maximize the Kepler Boost in the Overclocking Section.

One thing to keep in mind is the Kepler Boost is only throttled once per throttle point. For example, if the GPU got up to 75C, it would be throttled exactly one time at 70C by 13MHz; It would not be throttled continuously until the temperature dropped back below 70C.

How to find your max Kepler Boost: Warning: Spoiler! (Click to show)
Max Boost (or Maximum Boosted Frequency, abbreviated MBF):
This is the maximum frequency that the core will run at under full-load when the GPU is free of both power and heat constraints. By that i mean, the card is utilizing it's maximum Kepler Boost without any throttling. The Max Boost is found by adding the max Kepler Boost to the Boost clock (Max Kepler Boost + Boost Clock = Max Boost). This is the most significant number when talking about the overclock on your Kepler-based GPU. It's the ONLY number alone that gives us any indication of the performance of your GPU. So please, when describing your overclock, use THIS number, not the offset, not the Kepler Boost, not the Boost Clock, and certainly not the GPU clock.

How to find your Max Boost: Warning: Spoiler! (Click to show)
Power Percent and Power Target:
The Power Percent is the percent of TDP (Thermal design power) the card is running at. So if the card's TDP is rated at 170W and the card is running at a Power Percent of 117%, that means it's drawing 17% more watts of power than TDP, or 199W. The Power Target on the other hand, is a user defined number that tells the GPU how much over TDP it can go. For example, if you were to set the Power Target to 109% then the card would be able to run about 9% over TDP. However, there does seem to be some additional headroom involved, so realistically, a 109% Power Target would allow the card to go up to around 115% Power Percent. Like i mentioned above in the Kepler Boost section, Kepler Boost is also throttled down based on Power Percent. If the GPU is set at 100% Power Target, then the card will only be allowed to draw enough power to boost up to ~100% of TDP, anymore than that and the card will start to throttle the Kepler Boost value by 13Mhz increments in-order to stay below this number. If you were to then set the Power Target to a higher value, then the GPU could boost a little higher until it either reaches its maximum Kepler Boost or it reaches this new Power Target.

How to find your Power Percent: Warning: Spoiler! (Click to show)
GPU/Memory Clock Offset:
The GPU Clock Offset adds a static amount to the Boost Clock. A +50 Offset will add 50MHz to the boost clock, however, it will NOT increase your Kepler Boost value. The Memory Clock Offset simply increases the frequency of the memory. A +50 on the Memory Clock Offset will add only 25MHz (or one half of the offset value) to each memory module, or +100 (or double the offset value) to the effective memory value. So, if your base memory frequency is 1502MHz (or 6008 effective) then a +50 memory offset will increase the memory frequency to 1527MHz (or 6108 effective).

>>Link back to the table of content<<
Edited by SeanPoe - 6/3/12 at 1:28am

Software: Benchmarks and Overclocking Tools

Drivers
The first thing you will need to do is make sure you have the newest driver for your graphics card. Having the newest, most stable driver version will eliminate numerous false positive crashes that are solely related to drivers and not your overclock. The newest driver for the 670 can always be found here. If you're upgrading from an AMD graphics card or a really old Nvidia driver version, you might want to manually remove all traces of previous drivers before installing the new drivers. This is entirely optional but if you end up getting Red/Brown Screens, then the first thing you should try is doing a clean driver installation.

How to manually remove old AMD drivers: Warning: Spoiler! (Click to show)

How to manually remove old Nvidia drivers

Software and Benchmarks
The next thing you need to do is get various benchmarks and overclocking software downloaded, installed and setup.

Precision-X:
You have three choices in the software overclocking category: EVGA Precision-X, MSI AfterBurner, or Asus GPU Tweak. I've tried all three exhaustively and can conclusively say that Precision-X is the clear winner here. Asus GPU Tweak just didn't work at all. With the exact same settings that i used in Precision-X and Afterburner, GPU Tweak would crash my graphics drivers. Asus GPU Tweak has also caused Red Screens for a lot of people. Both Afterburner and GPU Tweak's voltage setting seems to be non-functional too. With Precision-X you can set the voltage slider to 1.175 and it actually pegs the voltage at 1.175 when the card is at a high Power Percent. This can make an otherwise unstable overclock stable. The only downside with Precision-X is its graphs don't have min/max values like the other two, but that's not really a big deal, especially if you have GPU-Z too.

Precision-X can be downloaded here, you will have to sign up on the EVGA website in-order to download it though.
I recommend that you uninstall both MSI Afterburner and Asus GPU Tweak and then restart your computer before installing Precision-X.

Precision-X Setup: (Click to show)
GPU-Z:
Download Link

I talked about this program a bit above in the Vocabulary section. If you haven't already, it's a good idea to get this program. It makes a great complement to Precision-X as you can use this to easily track your max values on the sensor tab and show your Boost Clock and Memory Clocks.

Unigine 'Heaven' DX11 Benchmark
Download Link

You will want to max out all of the settings here, so it looks like this:

You should set the Resolution to your monitors maximum supported and/or the resolution you plan to use in the majority of games. Using a lower resolution is fine for standardized benchmark comparisons, but for stress testing an overclock, using a lower resolution than you actually plan on using in games might change your card's maximum Power Percent under load, making your GPU unstable at more demanding resolutions.

Be sure that you remember to click the 'Benchmark' button in the top-left corner of the screen after the program loads to actually start the Benchmark. I'd also suggest turning off the sound by clicking the button on the top-left menu bar. Also, if you alt-tab out of Heaven, you will need to close it completely after that to get a real score. If you notice that your score suddenly dropped by 10 or more FPS, then your card is probably stuck in 2d mode. To fix this, first try restarting your computer.

To stop the benchmark at any time, press the esc key and then click quit.

When saving a score, be sure you leave the .HTML extension on the file name. If you accidentally remove the extension when you save a score, you can always go back and rename the score. Just add .HTML to the end. I also suggest that you use a naming convention that means something so you can easily compare scores from different runs. Something like 'CCC-MMM-A' where 'CCC' is your core offset value, 'MMM' is your memory offset value, and A is a rough estimate of how many artifacts you saw.

3DMark 11
Download Link

If you're using the free version then each time you open 3dMark11 you will need to click the "upgrade later" button. Switch the preset to 'Performance' if it's not already there and switch it to run only the Benchmark Tests in Centered Mode.

The esc key can be used at any point during the benchmark to end it prematurely.

Furmark:
Download Link

This is an old favorite of most GPU overclockers, but unfortunately, it has very little value in testing Kepler-based GPUs. The problem is it unrealistically increases the Power Percent to the absolute max the entire time forcing the GPU to throttle all the way down to the Boost Clock. But it still has a purpose, using a custom resolution we can use this program to simulate temperatures that you would likely see in extremely demanding games. This makes it easy to setup fan profiles. Other than that, it has no value.

I've found that setting it to a custom resolution of 228x228 with 8XMSAA realistically simulates what the most demanding games ever will produce in terms of temperature and power draw.

With these settings it will increase the Power Percent to the max, but just under the throttle point. It will also increase the GPU temperature to about 2C higher than even the most demanding maxed-out games will.

Other Benchmarks and games:
These are completely optional, but once you think you're stable in 3DMark11 and Heaven, it might be a good idea to test your overclock in a couple other games and benchmarks just to be sure. Some good free options are:
Alien Vs Predator Benchmark
S.T.A.L.K.E.R.: Call of Pripyat Benchmark
S.T.A.L.K.E.R.: Clear Sky benchmark
Mafia 2 Demo with Benchmark
Tom Clancy’s H.A.W.X. 2 Benchmark

>>Link back to the table of contents<<
Edited by SeanPoe - 6/3/12 at 11:22am

Overclocking:
Now that the new Vocabulary has been defined and you have all of the software and benchmarks installed and setup, it's time to move on to the main attraction
There's two methods for overclocking: the long method that will result in a nearly perfect overclock; and the short method, where you'll get pretty close but it won't be the absolute max. Either method is fine, it just comes down to how much time you want to devote to your overclock. I will be explaining both methods below.

The Long Method -- Estimated Time: 3 Hours
This method will get you to within about .25% of your card's maximum overclocked performance. Three hours is a lot of time, but you only need to do this one time and then your card is overclocked for the rest of it's lifespan. If you don't have the time or patience to devote three hours to this project, then skip down to the Quick Overclocking section of this guide.

Finding your Maximum Stable Core Clock

Step one: Preparation:

• Be sure your fan curve is setup like i mentioned above in the Precision-X setup section. It's important to keep your GPU under 70C at all times to prevent thermal throttling.
• Temporarily increase both the Power Target and Voltage slider to their maximum value. This will minimize the throttling caused by going over the Power Target.
• Go to the Nvidia Settings menu (you can open this from the taskbar), go to Manage 3d Settings, Global settings tab, scroll to the very bottom and make sure Vsync is set to off. You can turn this back on after you find your maximum overclock if you wish.
• Increase the GPU Clock Offset by whatever it takes to get your Boost Clock to 1100MHz. So that means if you're using a reference model with a default Boost Clock of 980 you will need to increase the offset to +120. If you're using a Factory overclocked card, then you will need to increase the offset by +10 to +50 (or until you hit 1100MHz Boost Clock) depending on which card you have, just use GPU-Z to keep track of your Boost Clock. If you have an Asus TOP 670 then i suggest down-clocking the card by 37MHz (-37 offset) for now. The reason being, someone Asus TOP models are factory overclocked to an unstable level.

Now double-click on one of the graphs in Precision-X to bring up the expanded Hardware Monitor (you may need to click the 'Performance Log' button before you can do this). Right-click on the expanded hardware monitor and then click 'clear history'. Now start a run of Heaven Benchmark. This will give us some information about your card's overclocking capability. As soon as the Benchmark finishes, save your score (name it CC-MMM, where CC is your core offset, and MMM is your memory offset, so here it should be 120-000 or something similar), hit esc, and then quit Heaven. If you didn't make it to the end of the benchmark for whatever reason (ex, the benchmark crashed, your display drivers crashed and stopped responding, or you had a BSOD or a Red screen) or if you saw graphical artifacts during the benchmark (Bright green/blue spots), then skip to Step 4. If you made it through the benchmark, then right-click on the Expanded Hardware Monitor and click pause. It should look something like this:

• Notice how the voltage is constant the entire time? That's because we increased the voltage to 1.175V in Precision-X which locks the voltage at the maximum during heavy load.
• Now mouse through the Power Percent line and try to fine the highest peak. The lower this peak is, the better. A low power percent at this stage usually indicates you have a lot of overclocking headroom left
• Now mouse through the GPU Clock. What you're looking for is a perfectly straight, constant line (some drops at the very beginning and end are normal). If the line isn't straight at this stage, then that's an indicator that your card is already being throttled and you have little to no more core overclocking headroom.
• You can go ahead and calculate your Kepler Boost at this moment if you wish. Just take your Max Boost Frequency and subtract your Boost Clock. The higher your Kepler Boost, usually the better your card will overclock. Keep in mind though, this isn't a hard rule; some people with low Kepler Boosts can still get high overclocks.
• Now mouse through the temperature line. You want to make sure that you didn't go over 70C. If you did, then refer to the troubleshooting section called "Reducing Temperatures to Eliminate Throttle" below. If your temperatures are below 70C, this is a good sign and means you can keep overclocking for now.

Step 2: Increasing your offset:

Increase the GPU Clock offset by an additional 20MHz, clear the Precision-X hardware monitor and unpause it, and start up a second Heaven Benchmark run. Immediately after the Heaven run completes, save your score (use a similar naming convention as you did above) and pause the Hardware Monitor. If you crashed or have any other symptoms of instability, go to Step 4. If your Heaven score decreased, that can mean one of three things. Either your overclock is unstable, your overclock is pushing you over the thermal or power throttle zone reducing your overall performance, or the benchmark was inconsistent. This is something that you need to be very vigilant of. Just because you hit a higher Max Boost frequency doesn't mean your GPU is performing better. Here's an example of what I'm talking about:

Compare this graph to the one i posted above in Step 1. In the first graph my GPU core was at a constant 1241MHz the entire time. As you can see here, my Max Boosted Frequency has gone up to 1264 but there's a ton of oscillation and it's also dipping extremely low too, even down to my base boost clock of 1120. So even though my card momentarily has a higher Max Boost, my overall performance has decreased substantially. This is why it's extremely important to always analyze the graphs after each offset increase. Higher is not always better with Kepler GPUs. If this is what your GPU Clock graph looks like (and/or even if you're seeing small 13mhz drops), go on to Step 3.

Here's an example of what a successful offset increase looks like:

Now you see my Max Boost is running at a constant 1264 with no dips. This is what you want to see with each additional increase in offset. So as long as your Heaven score hasn't decreased, and your Max Boost is running at a constant Frequency, then this means you're ready to increase the offset some more. Keep repeating this step (step 2) until you meet the criteria to go on to another step.

However, if your GPU clock doesn't show any variance, yet your Heaven score has decreased by more than 5 points, then i suggest that you re-run the Heaven Benchmark. If your score is still lower than the previous offset, then revert back to the previous Offset that had the highest score, and re-run Heaven again. If that score is still higher, then that suggests that your overclock with the increased offset is unstable. If this is the case, increase your offset back to the one that produced the lower score and then go to Step 4.

Step 3: Dealing with Frequency Oscillation:
If you're on this step then that means you're seeing your Max Boost Frequency constantly changing which will often times result in a performance decrease. First, check that your temperatures are still under 70C at all times. If they are, then you know it's not your temperatures causing the problems. If your temperatures are going over 70C, then refer to the Troubleshooting section below, "Reducing Temperatures to Eliminate Throttle."

Next, check to see if the Frequency drops also correspond with a GPU usage drop. If the GPU usage drops to less than around 90%, the GPU core clock might be throttled down just to save power but this won't lead to a performance decrease. Here's an example:

You see how each time my GPU usage drops my core clock also drops? This is because whenever there's a drop in GPU demand, for example when you're on a loading screen between sections in a benchmark or game, the core clock also drops in order to save power. This is completely normal and just means the GPU is throttling itself whenever additional core frequency is not beneficial, so you aren't actually losing performance. If this is what you're seeing, then go back to Step 2 and ignore the frequency oscillation that also corresponds with GPU usages drops too.


Now, check to see if your Power Percent is close to your set maximum Power Target (remember, Power Percent can go 3-10% over the set Power Target). So if your Power Target is set at 117%, and your Power Percent is at 120%-124%, then that is probably the reason for your oscillating frequencies. The only solution I've found to stabilize the variance in Max Boost caused by a maxed out Power Percent is to reduce the voltage slider. The reason this works is it reduces your Power Percent slightly but it might also cause instability. Try reducing the voltage slider back to default and check if your Max Boosted Frequency stabilize. Usually though, this method might stabilize your frequencies but it will cause instability too. If it causes instability, go to step 4. If not, and it fixed your oscillating frequencies, then go back to Step 2.

Step 4: Fixing Instability
So either your application crashed or produced an error, your display driver crashed or became unresponsive, you Blue-screened, you've notice graphical anomalies or your Benchmark score has decreased, or changing voltage doesn't stabilize the variance in your Max Boost. This usually means you're over your maximum stable offset. The only way to make your card stable at this point is to reduce the offset. The first thing you'll need to do is save a profile of your current settings in Precision-X (this is done by clicking the profile button at the bottom-left and then right-clicking a number), now close Precision-X and re-open it. The reason you have to restart Precision-X after each crash is because Precision-X will sometimes get bugged and not transfer your settings to the GPU. Restarting it is the only way to fix this. Now that you've restarted Precision-X, left-click on your saved profile and then click apply to load your previous settings, go to the voltage control and increase it to the max again (crashing resets voltage to default, and you can't save voltage in profiles). These underlined bits are extremely important. You will save yourself a lot of frustration later if just remember to restart Precision-X after each crash and then reset your voltage.

Now reduce your core offset by 5Mhz, clear the hardware monitor (and unpause it if applicable), and start a Heaven run. After the heaven run completes, save your score and pause the hardware monitor. If you still crashed or had other symptoms of instability, keep repeating this step until you reach stability and you reach a flat-lined (ie, constant) GPU Clock on the Precision-X graph. Be sure you restart Precision-X and reset your voltage after each crash. If you reach stability with a constant GPU Clock, go to Step 5.

Step 5: Fine Tuning
Now that you're within 4MHz of your maximum stable overclock, it's time to increase the offset in smaller steps. Increase the core offset by 1MHz, clear the hardware monitor (and unpause it if applicable), and start a Heaven run. After the heaven run completes, save your score and pause the hardware monitor. If this 1MHz increase caused instability again or the GPU clock to become variable (ie, not a constant line on the GPU clock graph), then reduce the offset by 1MHz and go to Step 6. If neither of these two scenarios happen, then repeat this step up four times and then go to Step 6.

Step 6: Checking for Instability in other Applications:
Heaven Benchmark will get your GPU stable in about 99% of other games, benchmarks and applications, but it's still a good idea to test your overclock in a couple other applications just to be on the safe side. I'd suggest starting with 3dMark11, then any of the other free game benchmarks i posted above, then boot up a couple games that you typically play (BF3 multiplayer is usually a good stability testing game if you have it). If your GPU show signs of instability in any of these games or benchmarks, then reduce your core offset by 1MHz until the problem goes away.

Now that you've tested your overclock in some other applications, it's safe to say that you've now found your maximum stable boosted frequency. I suggest that you save these settings in a Precision-X profile for future.

Finding your Maximum Stable Memory Clock:

Step 1: Increasing your Memory Offset:

• Revert your GPU Core offset to its default (+0) offset but leave the Power Target and voltage slider at whatever it took to stabilize your Core (most likely the maximum for both of these).
• Increase your Memory Offset to +100
• Start up a Heaven Benchmark run. Be extremely vigilant during the entire Benchmark run, you're looking for any graphical anomalies or artifacts. These usually look like green, blue, or white light shafts or dots. I've found it's usually much easier to spot these if you move six feet away from your monitor so you can have most of the screen in visual focus. If you see any artifacts, go ahead and hit ESC and then click 'Quit' in the top-left menu to end the Benchmark early; there's no point in letting the Benchmark finish if you already know it's unstable. After the Heaven run completes, save your score (name it CC-MMM, where CC is your core offset, and MMM is your memory offset, so here it should be 00-100). If you crashed, saw any graphical artifacts or have any other symptoms of instability, or if your score decreased by more than 5 points, go to Step 2. If not, repeat this step until you do and then go to Step 2.

Step 2: Fine Tuning your Memory Offset

• Reduce your Memory Offset by 25Mhz. For example, if you were at +600 in Step 1 and that started to produce graphical anomalies, you should be at +575 now.
• Start up a Heaven Benchmark run. Be sure you're still looking for graphical artifacts during the entire run. Once the benchmark finishes, save your score. If you crashed, saw any graphical artifacts or have any other symptoms of instability, or if your score decreased by more than 5 points, then repeat this step until the problem is resolved and then go to Step 3. Be sure that you restart Precision-X after each crash like i described in the Core Clock portion of this guide.

Step 3: Fine Tuning your Memory Offset

• Increase your memory Offset by an additional 5Mhz. For example, if you had to drop down to +525 Offset in Step 2 to eliminate artifacts, then you should be at +530 Offset now.
• Start up a Heaven Benchmark run. Be sure you're still looking for graphical artifacts during the entire run. Once the benchmark finishes, save your score. If you crashed, saw any graphical artifacts or have any other symptoms of instability, or if your score decreased by more than 5 points, then repeat this step until the problem is resolved and then go to Step 4. Be sure that you restart Precision-X after each crash like i described in the Core Clock portion of this guide.

Step 4: Fine Tuning your Memory Offset

• Decrease your memory Offset by 1Mhz. For example, if you were able to increase your Offset to +540 in Step 3, then you should be at +539 now.
• Start up a Heaven Benchmark run. Be sure you're still looking for graphical artifacts during the entire run. Once the benchmark finishes, save your score. If you crashed, saw any graphical artifacts or have any other symptoms of instability, or if your score decreased by more than 5 points, then repeat this step up to four times or until the problem is resolved and then go to Step 5. Be sure that you restart Precision-X after each crash like i described in the Core Clock portion of this guide.

Step 5: Checking your Memory for Instability in other Applications

• Like I mentioned in Step 6 of the Core overclocking section, it's a good idea to check for stability in other applications too. At the very least, check for stability on 3DMark11. If you see artifacts in any other application, reduce the Memory Offset by 1MHz until the problem goes away.
• Now you know your maximum memory frequency. I suggest that you save this to a profile number.

Finding your Maximum Stable Total Overclock
Now that you found your maximum core and memory overclock independently of each other, it's time to attempt to put them together. With some luck, you'll be able to put the two together and be stable right off the bat.

Step 1: Combing both Maximums
Set Precision-X to your maximum core offset and maximum memory offset that you found in the above steps. Be sure the voltage is set to whatever you needed to stabilize your core clock (for most people, this should be maxed). Be sure the Power Target is also set to the max. Now clear the hardware monitor (and unpause it if applicable), and start a Heaven run. After the heaven run completes, save your score (name it CC-MMM, where CC is your core offset, and MMM is your memory offset) and pause the hardware monitor. If you crashed, saw artifacts, or had any other symptoms of instability, then go on to Step 2a. If not, then you're done! You're at your true maximum GPU overclock. Go on to the Fan section below.

Step 2a: Explanation of Core and Memory balance
Most likely you crashed in Step 1 when you first combined both maximums. This is quite normal. From here you usually have three paths that you can take. Either you sacrifice some core frequency in-order to maintain your maximum memory frequency, you sacrifice some memory frequency to maintain your maximum core frequency, or you sacrifice a little of both. Neither option is better than another on the whole, it really depends on your specific GPU and how much you have to give up of one to maintain the other. Some GPU's barely have to give up any of one to maintain the other. Also, some applications benefit more from memory than core and vice versa. Generally speaking though, 1MHz on the core is worth about 4MHz on the memory in terms of performance, keep that ratio in mind as you reduce things.

Based on a lot of trial and error with three difference GTX 670's (Asus TOP, Gigabyte Windforce, and an EVGA reference model) I've found that reducing both core and memory together produces the best results. When i only reduced the core frequency enough to stabilize the GPU I'd still have memory related artifacts. When i only reduced the memory, I'd have to give up a huge amount of memory to stabilize the GPU; I'd have to trade something in the ball park of 25MHz on the memory for 1MHz on the core, which if you remember the 1-4 ratio above, is a very poor trade. My maximum boosted frequency was also getting throttled by 13MHz at certain points in 3dMark11 because i was hitting the maximum Power Percent when i only reduced the memory. But when i reduced both together, I've found that i can keep both the core and memory frequencies rather high while eliminating the memory related artifacts without seeing a Power Percent throttle in 3dMark11. Because of this empirical evidence, I'm going to suggest in this guide that you reduce both together. It's up to you if you want to try the other two methods, they might net you slightly better performance.

Step 2b: Balancing the Core and Memory
Close and then reopen Precision-X, re-enter your maximum offsets (or re-enter the Offsets you used in your previous run if you're repeating this step), Power Target, and voltage.
Reduce the core offset by 1MHz and the Memory Offset by 5MHz. After the Heaven run completes, save your score. Keep repeating this step until you become stable enough to finish Heaven (ignore artifacts for now) and then go to Step 3.

Step 3: Finding your optimal Balance Point
If you finished Heaven but still saw artifacts, then reduce the memory offset by an additional 5MHz until the artifacts go away. For every 5MHz you reduce the memory you can typically increase the core offset by an additional 1MHz. Be sure you don't increase the core offset higher than 1MHz below your maximum stable core offset. I've found that staying 1MHz below your maximum stable core offset allows you to increase the memory quite a bit more while still retaining stability without artifacts. Repeat this step until the artifacts are gone and you're still stable. If at some point you do become unstable again (if you crashed, remember to restart precision-X) reduce the core value by 1MHz but leave the memory where it was. If that is stable again, then continue the pattern until the artifacts are gone. Once the artifacts are gone and you can complete Heaven, go to the next step.

Here's an example of Step 3: Warning: Spoiler! (Click to show)
Step 4: Fine Tuning the Memory Offset
Leave the core offset where it is and increase the memory offset by 1. Repeat this step until you see artifacts or crash, in which case you'd reduce the memory offset by 1, or until you've repeated this step a total of four times.

Step 5: Checking for Instability in other Applications:
Now, just like above, you should test your combined overclock in other applications. I'd suggest starting with 3dMark11, then any of the other free game benchmarks i posted above, then boot up a couple games that you typically play (BF3 multiplayer is usually a good stability testing game if you have it).

• If you crash in any of these games or benchmarks, then reduce your core offset by 1MHz until the problem goes away. Just be sure the crash is not caused by the game itself before deciding to reduce things.
• If you see graphical artifacts, reduce the memory offset by 1 until it goes away.

You should now be at your highest performing stable overclock. Go on to the Fan section of this guide.

---

Quick Overclocking -- Estimated time: 30 minutes Warning: Spoiler! (Click to show)

---

Setting up a Quiet and/or Efficient Fan Curve:
There's a couple of ways to go about this. Either you setup a fan curve that maximizes performance or you setup a fan curve that's quiet but sacrifices a small amount of performance. If you have a 670 with non-reference cooling (specifically the Gigabyte Windforce, both Asus models, or the factory overclocked Galaxy model) you can easily get away with a very quiet fan profile without sacrificing performance. However, with the turbine-cooled reference models, you'll most likely have to choose between a quiet profile or maximum performance. If you have low ambient temperatures (both inside and out of your case) then you can probably get away with a performance profile without too much noise.

Before we begin though, I'd like to briefly go over the tools we're going to be using to perfect these fan curves.

Furmark Simulator:
Download Link

By using custom settings in Furmark, we can simulate the temperatures of an extremely demanding game. These are the settings you should be using for all the tests below:


Based on playing various demanding maxed-out games (The Witcher 2, Skyrim, BF3-Multiplayer, etc) I've tuned these settings to closely simulate the temperatures you'll see in these games but about 3C higher. This gives you some headroom if there's an especially demanding part of a game or benchmark that increases temperatures beyond normal levels.

Finding your idle Fan Speed:
Now you need to find your idle temperature. To do this, be sure you don't have any benchmarks or videos open that might pull the card out of idle. Also be sure your have your stable overclock settings loaded and applied.

First, set your fan to manually run at 25% by un-checking the Auto box and then moving the arrow on the side down until the fan read-out shows 25% and then click apply:

You will need to wait 1-3 minutes for temperatures to stabilize. This resulting temperature is is your idle temperature. Now set the fan back to auto by re-checking the box, and then click apply. The reason we have to do this is because the manually-set fan speed minimum is only 25% and you might need to go lower than that to find your true idle fan speed. Open up the fan-curve editor and set a point 3C higher than your observed idle temperature and at 25% fan-speed. This is to account for ambient temperature variance. Now keep the X-axis set at your idle+3C temperature and reduce the fan speed in 5% increments (be sure to click 'OK' after each change and wait 1-2 minutes for temperatures to stabilize), until you see the temperature of your GPU increase by 1C. Be sure you don't go below 10% though. Now add 2% to this number and this is the fan speed required to maintain your idle temperature. For me, my idle temperature is 30C (so i set the X-axis at 33C), and i observed that at 10% fanspeed my temperature rose to a constant 31C, so i increased the fan-speed by 2% to 12%, and that's what i need to set for the Y-axis:


The reason you want the fan speed to be constant at idle is so the noise profile is also constant. If you didn't have it at a constant speed, then when the fan has to increase even by as little as 1% it can change the noise profile slightly and with some louder fans (or fans that have frequent pitch changes) this can be noticeably annoying, especially when it oscillates by +/-1% constantly with each variance having a slightly different pitch.

Now that you've set your Idle fan speed, you'll need to decide if you want to go with a quiet fan profile that might sacrifice some performance or a maximum performance fan profile.

Maximum Performance: Warning: Spoiler! (Click to show)
Quiet Profile: Warning: Spoiler! (Click to show)
>>Link back to the table of contents<<
Edited by SeanPoe - 6/3/12 at 11:17pm

Troubleshooting:

How to Receive Help With Your Overclock in This Thread:
If you're having issues with your overclock and would like help, then please post in the following format:

Here's an example: Warning: Spoiler! (Click to show)

---

Reducing Temperatures to Eliminate Throttle: Warning: Spoiler! (Click to show)

---

Frequently Asked Questions:
Work in progress, will update as more questions are asked

>>Link back to the table of contents<<
Edited by SeanPoe - 6/2/12 at 9:57pm

Videos:
ASUS GTX 680 Direct CU II vs Reference: GPU Boost, Overclocking, Temps & More! by motherboardsorg
This video is about the 680 but a lot of the information is transferable to the 670 too. It's a great primer on how Power Percent and temperature can affect overclocks.

GTX 670 DirectCU II Top Graphics Card Performance Overview by asusrog
The title of this video is slightly misleading, it's actually an overclocking tutorial for the Asus 670 TOP.

Additional Threads:
[OFFICIAL] GTX 670 Overclocking
This thread is the 'Official' thread to brag in about your 670 overclock.

[Official] GTX670 Reviews Thread
The 'Official' 670 review thread with links to most all of the online 670 reviews.

[Official] NVIDIA GTX 670 Owners Club
The 'Official' 670 owners club. This is probably the best place to ask generalized questions about the 670.

Asus 670 DirectCU II TOP
Ignore the 'TOP' in the title, this thread is a place for all Asus 670 owners to ask question and get help from other Asus 670 owners.

Gigabyte GTX 670 OC Version *HANDS ON*
Ahh, the thread that started it all. This is one of the very first threads on OCN that had first-hand information about the 670. It has now turned into the unofficial Gigabyte 670 owners' thread of choice.

GTX 670 Availability
This is the thread to watch if you're looking to buy a 670. Even though it's died down a bit now, people still post here from time to time when certain online retailers get a new shipment of the 670's in.

Brettjv's 470 SLI vs 670 Bench-Off (both OC'd) ... (Now with 7970 OC'd!)
The title of this thread is rather misleading, it's more of a collection of benchmarks comparing SLI'd 470's, the 670, the 680, and the 7970.

Rock Heaven Your Hardest!!! (670/680 Users)
Once you get your card to its maximum overclock, you should head over to Brett's thread and post up your score! This is also a good place to see how your 670 stacks up to the rest.

Closing Note:
If anyone has any other suggestions for this guide feel free to post them here and i'll try my best to incorporate everything into the original post.

Hope you enjoyed the guide and good luck with your overclock!
Edited by SeanPoe - 6/3/12 at 1:10pm

Sorry for the formatting problems, need 3-5 more minutes. I will update the main post and title when i finish

Ready to go!
Edited by SeanPoe - 6/2/12 at 8:28pm

That took way longer than 10 minutes to fix all my formatting problems after i copied this from wordpad

It's done and ready now though