Single page Print

Overclocking
The process of overclocking a GeForce GTX 980 Ti isn't especially difficult since it all happens in software, but it's not terribly straightforward. Nvidia's GPU Boost algorithm factors in multiple variables in order to determine the best possible clock speed from moment to moment, and a number of things can limit clock speeds: temperatures, power limits, voltage peaks, frequency caps, and so on. Current GeForce cards come with two rated clock speeds: a "base" clock that is a guaranteed minimum and a "boost" clock that represents a typical speed while gaming. (Yep, the boost clock is not a maximum. The GPU can range higher than that under the right conditions.)

As you might be gathering, actual GPU clock speeds will vary depending on the workload. Before overclocking these cards, I decided to establish a baseline by seeing what speeds they run in a very demanding program: MSI's Kombustor, specifically the "Furry PQTorus" test based on FurMark. This application represents a worst-case thermal scenario for modern GPUs and pushes things much harder than any game I've seen. We've used Kombustor in past reviews for this very same purpose.

Here's how each of the cards handled in FurMark.

  GPU
base
clock
(MHz)
GPU
boost
clock
(MHz)
Memory
speed
(GT/s)
Kombustor
GPU
voltage
Kombustor
GPU
clock
(MHz)
Reference GTX 980 Ti 1000 1076 7000 1.018 1189
Asus Strix GTX 980 Ti OC 1216 1317 7200 1.030 1240
EVGA GTX 980 Ti SC 1102 1190 7000 0.999 1050
Gigabyte GTX 980 Ti Gaming G1 1152 1241 7000 1.003 1177
MSI Gaming GTX 980 Ti 6G 1178 1279 7100 1.043 987

Right away, you can see that things get a little weird when we push these big GPUs with an atypical and very intensive workload. The delivered clock speeds and voltages vary widely.  Only the reference card, with the tamest tuning of the bunch, runs at a speed higher than its boost clock. You'll know from the previous page that these numbers aren't a very good predictor of in-game performance. I'll probably want to switch to a more typical workload next time I decide to overclock a batch of big GPUs.

Regardless, Kombustor is still a pretty good stress test, if nothing else. I was able to use it for that purpose while seeking the max possible clock speeds for each of the cards.

I decided to use EVGA's Precision X tool for overclocking this time around. Here's what I was able to achieve by prodding each of these cards for a while. Of course, since this overclocking and is in no way guaranteed, your mileage may vary. The peak possible clock speed at a given voltage will vary from chip to chip.

  GPU
clock
offset
(MHz)
Boost +
offset
(MHz)
Memory
speed
(MT/s)
Kombustor
GPU
voltage
Kombustor
GPU
clock
(MHz)
Kombustor
GPU
temp. (°C)
Reference GTX 980 Ti +270 1346 8000 1.012 1101 84
Asus Strix GTX 980 Ti OC +60 1377 8000 1.055 1328 83
EVGA GTX 980 Ti SC +140 1330 8000 0.999 1177 77
Gigabyte GTX 980 Ti Gaming G1 +120 1361 8000 1.074 1354 74
MSI Gaming GTX 980 Ti 6G +60 1339 8000 1.043 1215 78

The first thing you might notice in the table above is that third column, "boost + offset." That's effectively the peak speed I was able to wring out of each product, and there's not much distance between the slowest and fastest options. Cards like the MSI with higher default clocks would only tolerate smaller offsets, while conservatively tuned boards like the reference model handled larger offsets well. At the end of the day, GM200 chips tend to have similar limits, regardless of which card they're on.

One reason they're so similar is that Nvidia has limited the peak voltage. Each and every one of these cards will allow you to raise the default GPU voltage by 87 mV and no more. I simply took each card to its max when I tested. I'd need more voltage to push the GPU offsets any higher without stability problems.

As for memory speeds, well, I generally pushed from the default of 7Gbps to 8Gbps and stopped. Each of the cards was able to handle that speed, but things got dicey in a hurry when I pushed past that mark. Since a number of GDDR5 chips are involved on each board, what you can get on the hairy edge of memory speeds will likely vary even more from one board to the next.

When overclocked using the settings above, all of the 980 Ti cards limited their clock speeds in Kombustor for the same reason: hitting their power limits. The reference board briefly hit a temperature limit instead, until I raised its temperature tolerance by a few degrees. Then they were all power limited. You can raise the power limit a bit in Precision or similar utilities, and I maxed it out for each contestant. Most cards topped out at 110% of their default, but the MSI allowed a 120% setting. The Gigabyte went to 130%. Regardless, power was always the limiting factor for clock speeds in Kombustor.

I should say that it's entirely possible that my simple formula of using the peak voltage with the peak clock offset might not be the most optimal config in every case. One could possibly extract more performance from a power-limited GPU by keeping its voltage lower. But finding that magic sweet spot involves serious tweaking and tedium, and I just didn't have the extra hours to dedicate to that process for every card. At least next time, I'll use something else generate the GPU load, so we can see more typical clock behaviors.

Of course, the true test is how these overclocked configurations perform in games.

Turning up the clock speeds and voltage/power limits for these 980 Ti cards adds another 10% or so to their performance. Asus and Gigabyte are once again at the top of the pack when overclocked.

There's a reason the Asus and Gigabyte cards are able to achieve the highest stable boost + offset clock frequencies and thus the highest frame rates: they use more power than everything else in order to stay stable. I'm not sure whether they're secretly throwing more voltage at the GPU or what, but the difference in power draw versus MSI and EVGA is about 18W at stock speeds and ~30-50W when overclocked.

Even when topping the performance charts, the Asus Strix is among the quietest cards while overclocked. That's a nice combination of attributes. The Gigabyte card achieves similar performance, but it makes a more noise in its efforts to keep GPU temperatures lower.