|MSI GTX 660-Ti Power Edition Video Card|
|Reviews - Featured Reviews: Video Cards|
|Written by David Ramsey|
|Thursday, 16 August 2012|
Page 15 of 17
MSI N660Ti PE Overclocking
Overclocking Kepler-based cards isn't the way it was in "the old days". With previous generations of GPUs, you overclocked simply by increasing the clock frequency (and perhaps power and voltage) until the card failed. Although some cards could "protect themselves" by throttling power draw and frequency at the high end, there was no unified strategy for managing top-tier performance.
All that changes with Kepler. Kepler GPUs will vary their clock speeds from "somnolent" to "insane" depending on a number of factors, and your overclocking settings are treated as requests rather than commands. NVIDIA uses both hardware and software constraints to achieve the maximum performance possible under a given set of parameters, taking into account power draw and temperature as well as (according to NVIDIA) "many other" factors-- and no, they don't want to talk about what those other factors are.
This means that when you see NVIDIA or other vendors talking about "boost clock", don't assume that this clock speed is what a Kepler-based card will run at under load. According to NVIDIA, the "boost clock" is the average frequency the GPU runs at under NVIDIA's test suite. MSI says the boost clock for the N660Ti Power Edition is "1097MHz", which means that they guarantee the card will average this speed under load. In some cases it may be faster. In some cases it may be slower.
NVIDIA's performance strategy means that identical Kepler-based cards from the same manufacturer are not guaranteed to have identical performance-- in fact they almost certainly won't. While they'll boost to at least the quoted boost clock under load, how much further they will go will vary from card to card.
With this in mind, I used MSI's own Afterburner 2.2.3 to increase the voltage and power draw to that maximum allowed for this card (+100mv and 114%, respectively), and then set out to achieve the highest GPU clock and memory clock offsets I could. This turned out to be 160Mhz and 470MHz, respectively.
In the benchmarks, this resulted in a maximum observed boost clock of 1,280MHz. Again, remember that this represents the maximum I saw, and is not necessarily the average clock speed across all of the benchmarks. Yes, it would be nice were there a utility that actually would record average GPU clock speeds on the fly.
This overclock propelled the N660Ti Power Edition into rarefied territory, enabling it to beat a reference design GTX680 in some tests. Below I summarize the performance increase overclocking yielded in each benchmark (in those that were run at multiple resolutions, I include results only from the highest resolution):
NVIDIA claims that although cards based on 660 series GPUs are priced to compete with 7800 series AMD Radeons, the performance can "often" compete with a Radeon 7950. The XFX Radeon 7950 I had for comparison is a factory overclocked card, but let's check out that claim anyway. For this chart I'm comparing both cards with their as-delivered factory overclocks:
So across this particular set of benchmarks the MSI N660Ti Power Edition is 3.92% faster than the XFX Radeon 7950 Black Edition Double Dissipation. While this number doesn't really mean anything in absolute terms-- note how the differences tend to be rather large (>20%) up or down depending on the benchmark-- it's still amazing that this card can credibly be said to be in roughly the same performance arena as the much more expensive 7950.
Interesting note: when using Afterburner 2.2.3 to overclock MSI's N680GTX Lightning, I found that I had to constantly watch the voltage offset, since the card would tend to move it down. I never saw that happen with the N660Ti.
I'll present my final thoughts on this card in the next section.