ASUS GeForce GTX 660 DirectCU II Overclocking Review
We put our new ASUS GeForce GTX 660 through the ringer of overclocking and make real world gaming comparisons. If you are thinking the new GTX 660 (GK106) GPU will be a good overclocker like its bigger brother GK104, you may be in for a surprise that puts the new GTX 660 in a new light.
On Thursday, September 13th NVIDIA launched the brand new GeForce GTX 660 GPU. We evaluated a full custom retail ASUS GeForce GTX 660 DirectCU II video card. This video card was not the OC or TOP version, but based on the reference clock speeds of the new GTX 660. We found better than expected performance out of the GeForce GTX 660, but it has some obstacles to climb against the stiff competition thanks to AMD price drops that have recently taken place. One area we were not able to cover in the original article but hoping would be the saving grace of the GTX 660 was overclocking. In this article today we will look at overclocking, which may surprise you compared to other "Kepler" architecture GPUs such as the GTX 660 Ti, GTX 670, and GTX 680 when it comes to overclocking potential.
First let's talk about what we have, and then we will talk about some limitations in the GeForce GTX 660 that hamper overclocking. The GeForce GTX 660 is using the same "Kepler" architecture as the GTX 660 Ti, GTX 670, and GTX 680. However, the GTX 660 is using new silicon; in this case it is called GK106., whereas the GTX 660 Ti, GTX 670 and GTX 680 all use the GK104 silicon. The new GK106 is still manufactured on a mature TSMC 28nm manufacturing process, and contains 5 SMX units. With a mature 28nm process, a newer chip, and less CUDA cores, you'd think overclocking would be at least equally as good as the more power hungry GK104 chips. Well, this is not really the case as we will explain below. There are some limitations that are holding back the potential for overclocking GTX 660 video cards.
ASUS GTX 660 DirectCU II Overclocking
The ASUS GTX 660 DirectCU II video card we have is as mentioned running at the stock frequencies of a GTX 660. That is it is running at 980MHz base clock and 1033MHz Boost clock. We've experienced the DirectCU II heatsink and fan keeping the temperatures very cool on this video card, so temperature is not a factor in the subpar overclocking in our opinion. In our testing, we found that with the video card’s in-game frequency was running at was 1084MHz. So this means that the video card already runs over the Boost clock of 1033MHz; this is 51MHz over the Boost clock of the video card, or a 5% clock increase. This is important to keep in mind as we discover below the Power Target option we have. Also keep this in mind, the TDP of the video card is 150W board power, this is set and capped by NVIDIA in the Power Target.
You can see above what the default settings are in GPU Tweak and GPU-Z, the base clock and Boost clock and fan speed and voltage is all on target. We found that the voltage stays at 1.162v during gaming with the card straight out of the box. At 1.162v it runs again at 1084MHz in games and the frequency is stable, meaning it doesn't vary or change while gaming, even in power hungry game scenarios.
In the screenshot above we are showing the maximum Power Target the GTX 660 allows, 10% (110 on the slider over 100) from the base. This is by far the lowest Power Target percentage of any Kepler generation GPU to date. On the higher-end GK104 GPUs the Power Target can be set to +20% and here we are only able to set +10%. We already explained above that at 1084MHz it is operating 5% above the Boost clock, already eating up some of our Power Target room. With only +10% Power Target, we aren't going to have the power needed to allow stable high overclocks, and our tests proved that. We have data that outlines all this.
Before we dive into the data, let's also talk about the voltage. Out-of-the-box the voltage is 1.162v, and we can set the slider up to 1.175v. What this means is that if we set it to 1.175 the voltage will never drop below that value, but, GPU Boost can also utilize this voltage on its own without our help. Increasing voltage to the overall GPU and board power of a video card makes that value rise. We are already dealing with a low TDP cap on this video card, and even with a +10% increase, increasing the voltage takes us into power territories that retard the useable overclock. Let’s explain.
GPU Boost and Board Power Connection
The way NVIDIA's varying GPU Boost works is that if the board power ceiling is not reached, the GPU clock speed will go up until it reaches the ceiling. If you go over the total board power, either with voltage, clock speed, or both, the GPU will downclock itself to keep the TDP in check. Think of it akin to a governor on a car. In this case, the Power Target is the governor, and we can only set it up +10%. Therefore if you go over the board power, the GPU clock speed is actually dynamically vary to keep power in check and has the potential to actually underclock below the 1084MHz the card works at stock settings. In this case, your performance gets worse. We have data to show this.
Highest GPU Potential Overclock Without Crashing
Ok, let's start out by first telling you the clock speed we achieved successfully. We were able to get the actual in-game clock speed of this GPU up to 1170MHz Boost clock which resulted in a 1215MHz actual in-game frequency. Anything above that, from 1225MHz and up, instantly locked up during gaming. 1215MHz actual in-game frequency this was the sweet spot for stability of the GPU. But not the sweet spot for performance.
The problem at 1215MHz is that the actual in-game frequency jumped all over the place while we were gaming, and in many situations dropped below the 1084MHz the card is specified for. In fact, we saw GPU clock go all the way down to 888MHz in order to keep the TDP in check. This was with +10% on the Power Target. This inherently gave us worse gaming performance than the card started at; it was underclocking the GPU.
Here's what we had to do, we had to lower the GPU frequency until we found a value where the varying frequency in-game never dropped below 1084MHz. In other words, we had to lower the frequency until it stabilized while we were gaming and maintained an overclock and never became detrimental to performance. It was a painstaking process of manually adjusting downwards the Boost clock in GPU Tweak until we found the sweet spot.
Highest Stable In-Game GPU Frequency Overclock
We found the sweet spot to be all the way down to 1100MHz Boost clock, which operated at an actual in-game frequency of 1152MHz. In fact, even at this lowered clock rate the frequency still bounced around in Battlefield 3. The frequency went between 1125MHz-1139MHz-1152MHz. All of these frequencies were above 1084MHz, and thus an overclock giving us more performance. If we were to have raised the clock speed anymore, then on the low-end of this varying frequency it would drop below 1084MHz and thus give us worse overall performance. Therefore, to get the most stable overclock that maintained frequencies above 1084MHz, we had to stick with 1100MHz GPU Boost, which ran at 1152MHz in-game at the highest. We have data that shows that a clock setting of 1152MHz is faster than 1215MHz.
Overclocking the memory frequency also raises board power, and as we've found out raising board power on the GTX 660 is a bad thing for the overclocking enthusiast as we don't have much power to work with. Overclocking the memory alone we were able to get it from 6GHz up to 6.7GHz for an incredible memory overclock. However, when we did this, it actually hurt our GPU overclock. What happened was that the TDP started to get pegged, and the subsequent varying GPU frequency dropped lower than 1084MHz with such a high memory overclock. The memory overclock was affecting the overall board power hurting the GPU overclock. We had to back down on the memory overclock to a point where it didn't affect the GPU clock speed. Again, this was a painstaking task of manually downclocking the memory until the GPU frequency in-game stabilized above 1084MHz. We found this sweet spot to be 6.4GHz. So the memory is still overclocked, but not to its maximum potential due to the card’s TDP cap.
Voltage Affecting Board Power
Now we bring this back around to voltage. We also found it best to just leave the voltage slider alone on the GTX 660. If we raised it to 1.175v all it did was raise the board power and kill our GPU overclock. By leaving it alone, and letting it operate at 1.162v, board power went down and we were able to get that 1152MHz overclock. The same also goes for fan speed; yes even fan speed can affect board power. We had to leave it at "Auto" so that it didn't increase demand on power at all instead of maxing the fan speed out like we would normally do in overclock testing scenarios.
GTX 660 Overclocking Focus
The key to GTX 660 overclocking is to keep board power down as much as possible; this will help you get the best GPU overclock. The reason why we have to do this is because of the measly +10% we have allowable with Power Target. If we had +20% we'd no doubt see overclocks up to at least 1.2GHz on this GPU, but the low Power Target is keeping the overclocking at minimal levels.