Overclocking, Efficiency, Heat, And Conclusion
Overclocking and Stability
Surprisingly, our water-cooled Founders Edition card made it all the way to 2.1 GHz stably. It's possible that Nvidia sent us a golden sample, so there's no way to know if our result is representative of the cards available to everyone. GPU Boost takes the clock rates down by one notch as soon as the sensor hits 40°C. Consequently, we configured our fans and pump to keep the card under 45°C at all times.
To start, we looked at how high the GeForce GTX 1080 Ti Founders Edition would jump at its stock settings. Then, we slowly increased its base frequency, eventually leveling off at a GPU Boost clock rate of 2.1 GHz at 30°C. Ideally, we'd hit the same frequency at 40-44°C using more conservative settings, so long as nothing else throttles the GPU.
We're fortunate enough to have a version of MSI’s Afterburner utility unlocked especially for us. If you want access to similar settings before a new version of Afterburner is released, you can manually add your 1080 Ti to the third-party database using the VDDC_Generic_Detection entry under the VEN_10DE&DEV_1B06&SUBSYS_120F10DE&REV_?? key. A quick search online should turn up plenty of in-depth instructions on how to do this.
Despite setting the power limit to 120%, it didn't take long for us to run into Nvidia’s built-in limiter. Overclocking is pointless if it doesn't result in a stable and consistent experience. That's why we use an hour-long combination of Metro: Last Light at 4K, The Witcher 3 at QHD, and Ghost Recon Wildlands at Ultra HD. We further pushed stability by running rendering tasks using 3ds Max, Nvidia Iray, LuxRender, and FurMark’s one-hour stress test.
A quick glance at the power target display reveals 112%. This poses some questions that need to be answered, since neither the GPU frequency nor the power consumption was stable. Usually, the overlays applied by diagnostic utilities (like the one pictured above) are only refreshed once per second. That's too long of an interval; as we know, a lot can happen inside of one second.
Realized Clock Frequency and Power Consumption
As an example, we're using Metro: Last Light at Ultra HD because it presents a fairly consistent load profile. The Witcher 3 gave us some peaks that exceeded the ones we saw under Metro: Last Light, but its average load during longer gaming sessions was lower. Furthermore, Metro's scripted benchmark heats up graphics cards more than most other sequences.
Our GeForce GTX 1080 Ti does hit 2101 MHz during the Metro: Last Light and The Witcher 3 tests. However, GPU Boost clock rates fluctuate depending on load, dropping to 2088 MHz fairly often. The frequency even falls to 2 GHz on occasion, only to quickly recover.
The recorded power limit curves tell us that our 120% setting is just a maximum. It's never actually hit. There are brief excursions up to 119% that last for fractions of a second, but the limit immediately retreats back to 110-112%. The average hovers right around 114 to 115%.
This result corresponds with our average power consumption result, which lands at 287W across all eight test runs. We used an intelligent low-pass filter and a special form of resampling to keep the amount of data generated by our measurements manageable (as many as 6000 values per second would have quickly overwhelmed Excel).
A higher-resolution graph of one test loop shows just how frantically GPU Boost tries to intervene and limit power consumption. The simplified summary in the graph gives us an average of 287.1W, while the more detailed one below, without resampling, yields 286.9W. The two results are practically identical; the difference is within the measurement's margin of error, proving our filtered data is accurate.
Some plausible peaks do break through the 300W barrier, but the average gaming power consumption lands just shy of 290W. Measurements in excess of 300W on a card that hasn't been modified are quite simply false. Nvidia's protections make sure of this.
MORE: Best CPU Cooling
MORE: How To Choose A CPU Cooler
MORE: All Cooling Content
Efficiency at Different Clock Frequencies
Let’s take a look at The Witcher 3. Several readers have asked us to use a game other than Metro: Last Light, so here we go. After a bit of experimentation, we found that a higher frame rate at 2560x1440 produces a more consistent load than hammering the card at 3840x2160 in this title, so that's what we're using.
To evaluate efficiency, we increased the clock rate in 100 MHz increments along the achievable GPU Boost frequency curve. Since GPUs fall into different quality categories, we didn’t manually undervolt, but rather set a plausible base frequency and then lowered the power target until we hit a sweet spot for consistent operation.
We completed five runs for each frequency step, and discarded the best and worst outcomes. Averaging data from the remaining three runs gave us a final result for that clock rate. Naturally, the power consumption curve gets steeper with increasing frequency, whereas the curves for the average and minimum FPS show the opposite pattern. Interestingly, this trend is more pronounced for the minimum than for the average frame rate.
Impressive speed-ups are available almost all the way to GP102's maximum clock rate, even if performance doesn’t scale perfectly. In other words, more consistent GPU Boost frequencies would translate to even higher real-world FPS. Seeing that this would require a shunt mod (like the one we've seen applied to Nvidia's Titan X), gunning for even more performance isn't a trivial matter. Our German lab only has one GeForce GTX 1080 Ti right now, so we're not risking it at this point.
Temperatures and Infrared Measurements
High GPU Boost frequencies like the ones we're seeing require GPU temperatures in the 40°C range. The only way to achieve that is by using a open-loop water-cooling setup. Our solution produced a delta of just 7°C between the temperature of the GPU and the water exiting its block. Consequently, we skipped Aquacomputer’s active backplate. The measurements were taken after a 30-minute warm-up phase in our closed bench table.
The results are great: even with Nvidia's GeForce GTX 1080 Ti Founders Edition running at its peak overclock, the GPU's temperature never exceeds 44°C.
| Peak Temperatures | |
|---|---|
| GPU Diode | 44°C (GPU-Z) |
| GPU Package | 43.7°C (Infrared Camera) |
| Water (In) | 28.2°C (Sensor in Fitting) |
| Water (Out) | 36.7°C (Sensor in Fitting) |
| VRM Hot-Spot | 55.6°C (Infrared Camera) |
| Memory Block | 52.3°C (Infrared Camera, Area with Highest Temperature) |
| Ambient Temperature | 22.1°C (Infrared Camera, Reference Measurement Area) |
Our thermal readings, which come from Optris' PI640 infrared camera, document the solid copper block's enormous cooling performance.
A comparison to the stock air cooler is just sad.
Summary and Conclusion
The GeForce GTX 1080 Ti Founders Edition, just like Nvidia's Titan X (Pascal) before it, is a monster of a graphics card when you augment its cooling. On air, the board doesn't really make it past 1.9 GHz consistently, and even that requires you pay a price in extremely high noise levels. Enthusiasts spending $700 for high-end graphics might want to consider investing in water-cooling as well to realize the card's full potential.
A water block does cost upwards of $100 per card, but the rest of the loop can last for many years if it’s maintained well. Complete kits are available for as little as $300 (without the graphics card block), which makes sense as a one-time expense. If you have the spare cash, we think you'll enjoy the results. Otherwise you'll probably want to buy a good set of non-vented headphones instead.
MORE: Best Graphics Cards
MORE: Desktop GPU Performance Hierarchy Table
MORE: All Graphics Content
There's a budget GeForce GPU selling in China that not even Nvidia knew it made — RTX 4010 turns out to be a modified RTX A400 workstation GPU
US to patch loopholes that allow China to buy banned AI GPUs from other countries — new regulations include national quotas on GPU exports and a global licensing system
