Here’s the part that Nvidia doesn’t allow its board partners to pursue, but still offers up to gamers through utilities like MSI Afterburner: manual overclocking.
There is one slider you don't get access to, though. Voltage manipulation is locked out entirely. Other Pascal-based graphics cards can be tuned through small voltage increases. However, to protect GeForce GTX 1080, the GeForce GTX 1070 Ti is deliberately handicapped in this one area.
MSI's GeForce GTX 1070 Ti Titanium has a theoretical power limit of 239W. Our goal is to see how far the default fan curve takes us without making the noise level intolerable. Conversely, we also want to figure out how high the GPU Boost clock goes if we let the fan spin as fast as it can.
All of the board partner cards our German lab received overclock similarly. However, the cards with a maximum power limit of 120% are at a significant disadvantage because 2050 MHz requires their fans to spin as fast as they'll go.
|Graphics Card||Power Target||Fan||GPU Boost Frequency(After 30 Minutes)|
|Nvidia GeForce GTX 1070 Ti FE||120%||100%||2063 MHz|
|MSI GTX 1070 Ti Titanium 8G||133%||Auto||2050 MHz|
|Gigabyte GTX 1070 Ti G1 Gaming||120%||100%||2063 MHz|
|Gainward GTX 1070 Ti Phoenix GS||120%||100%||2050 MHz|
|iGame GTX 1070 Ti Vulcan X Top||134%||Auto||2063 MHz|
The resulting curves show how GPU Boost frequencies drop as temperatures rise.
The corresponding voltages show us that GeForce GTX 1070 Ti isn't held back by its power limit. In fact, an additional 6 to 8W wouldn’t have been a problem. Instead, voltage is the limiting factor.
Ultimately, an almost constant 1.05V pushes GP104 to an impressive 2126 MHz for a short period of time with the fan speed set to 100% (and while the GPU’s temperature is below 42°C). This is exactly one step below Nvidia’s maximum voltage of 1.062V.
Infrared Picture Analysis
We also want to know how MSI deals with the waste heat generated by an additional 50W of thermal energy. To this end, we compare the card’s stock fan curve to its operation at 100% duty cycle.
Even when those 10cm fans turn as fast as possible, our equipment only registers ~43 dB(A). That's less noise than many graphics cards under load in their stock configurations.
At 2050 MHz, we measure 68°C, which won’t be a problem for the board’s longevity. An 85°C reading from the memory is 10°C below Micron's specified maximum, while a 95°C peak above the VRMs is merely acceptable. In the long run, that number shouldn't go up any further.
Pushing the fans to their limit has a significant impact on thermal readings. Long-term operation at 2.1 GHz shouldn’t be a problem, since temperatures won’t be an obstacle.
The Pitfalls of Overclocking
So, what about some performance data? In short, it's not consistent enough for us to graph.
Our German and U.S. labs deliberately compared the same models and found that chip quality was dramatically different between them. Dialing in a 133% Power Target and +220 MHz GPU offset caused crashes on one board and not the other. In the end, our less impressive sample only allowed a +150 MHz offset, resulting in a ~2 GHz GPU Boost frequency with the fans running at 100%. This delta is large enough to cast doubt on any attempt at universal benchmark results.
Memory can bottleneck overclocking results too, even with a +150 MHz increase. Depending on the game and resolution, you might see 8%-higher frame rates at 2050 MHz, while in other situations you realize gains between 2-5%.
We simply need more samples before judging the overclocking qualities of GeForce GTX 1070 Ti with blanket statements. Some of you will get lucky, while others strike out. We're trying to manage expectations after running into our own inconsistencies.
MORE: Best Graphics Cards
MORE: All Graphics Content