Overclocking

For our final evaluation of the GTX 1080 and GTX 1070 Founders Edition cards, let’s take a look a overclocking.

Whenever I review an NVIDIA reference card, I feel it’s important to point out that while NVIDIA supports overclocking – why else would they include fine-grained controls like GPU Boost 3.0 – they have taken a hard stance against true overvolting. Overvolting is limited to NVIDIA’s built in overvoltage function, which isn’t so much a voltage control as it is the ability to unlock 1-2 more boost bins and their associated voltages. Meanwhile TDP controls are limited to whatever value NVIDIA believes is safe for that model card, which can vary depending on its GPU and its power delivery design.

For GTX 1080FE and its 5+1 power design, we have a 120% TDP limit, which translates to an absolute maximum TDP of 216W. As for GTX 1070FE and its 4+1 design, this is reduced to a 112% TDP limit, or 168W. Both cards can be “overvolted” to 1.093v, which represents 1 boost bin. As such the maximum clockspeed with NVIDIA’s stock programming is 1911MHz.

GeForce GTX 1080FE Overclocking
  Stock Overclocked
Core Clock 1607MHz 1807MHz
Boost Clock 1734MHz 1934MHz
Max Boost Clock 1898MHz 2088MHz
Memory Clock 10Gbps 11Gbps
Max Voltage 1.062v 1.093v

 

GeForce GTX 1070FE Overclocking
  Stock Overclocked
Core Clock 1506MHz 1681MHz
Boost Clock 1683MHz 1858MHz
Max Boost Clock 1898MHz 2062MHz
Memory Clock 8Gbps 8.8Gbps
Max Voltage 1.062v 1.093v

Both cards ended up overclocking by similar amounts. We were able to take the GTX 1080FE another 200MHz (+12% boost) on the GPU, and another 1Gbps (+10%) on the memory clock. The GTX 1070 could be pushed another 175MHz (+10% boost) on the GPU, while memory could go another 800Mbps (+10%) to 8.8Gbps.

Both of these are respectable overclocks, but compared to Maxwell 2 where our reference cards could do 20-25%, these aren’t nearly as extreme. Given NVIDIA’s comments on the 16nm FinFET voltage/frequency curve being steeper than 28nm, this could be first-hand evidence of that. It also indicates that NVIDIA has pushed GP104 closer to its limit, though that could easily be a consequence of the curve.

Given that this is our first look at Pascal, before diving into overall performance, let’s first take a look at an overclocking breakdown. NVIDIA offers 4 knobs to adjust when overclocking: overvolting (unlocking additional boost bins), increasing the power/temperature limits, the memory clock, and the GPU clock. Though all 4 will be adjusted for a final overclock, it’s often helpful to see whether it’s GPU overclocking or memory overclocking that delivers the greater impact, especially as it can highlight where the performance bottlenecks are on a card.

To examine this, we’ve gone ahead and benchmarked the GTX 1080 4 times: once with overvolting and increased power/temp limits (to serve as a baseline), once with the memory overclocked added, once with GPU overclock added, and finally with both the GPU and memory overclocks added.

GeForce GTX 1080FE Overclocking Breakdown

GeForce GTX 1080 Overclocking Performance
  Power/Temp Limit (+20%) Core (+12%) Memory (+10%) Cumulative
Tomb Raider
+3%
+4%
+1%
+10%
Ashes
+1%
+9%
+1%
+10%
Crysis 3
+4%
+4%
+2%
+11%
The Witcher 3
+2%
+6%
+3%
+10%
Grand Theft Auto V
+1%
+4%
+2%
+8%

Across all 5 games, the results are clear and consistent: GPU overclocking contributes more to performance than memory overclocking. To be sure, both contribute, but even after compensating for the fact that the GPU overclock was a bit greater than the memory overclock (12% vs 10%), we still end up with the GPU more clearly contributing. Though I am a bit surprised that increasing the power/temperature limit didn't have more of an effect.

OC: Rise of the Tomb Raider - 3840x2160 - Very High Quality (DX11)

OC: Ashes of the Singularity - 3840x2160 - Extreme Quality (DX12)

OC: Crysis 3 - 3840x2160 - Very High Quality + FXAA

OC: The Witcher 3 - 3840x2160 - Ultra Quality (No Hairworks)

OC: Grand Theft Auto V - 3840x2160 - Very High Quality

OC:  Grand Theft Auto V - 99th Percentile Framerate - 3840x2160 - Very High Quality

Overall we’re looking at an 8%-10% increase in performance from overclocking. It’s enough to further stretch the GTX 1080FE and GTX 1070FE’s leads, but it won’t radically alter performance.

OC: Load Power Consumption - Crysis 3

OC: Load Power Consumption - FurMark

OC: Load GPU Temperature - Crysis 3

OC: Load GPU Temperature - FurMark

OC: Load Noise Levels - Crysis 3

OC: Load Noise Levels - FurMark

Finally, let’s see the cost of overclocking in terms of power, temperature, and noise. For the GTX 1080FE, the power cost at the wall proves to be rather significant. An 11% Crysis 3 performance increase translates into a 60W increase in power consumption at the wall, essentially moving GTX 1080FE into the neighborhood of NVIDIA’s 250W cards like the GTX 980 Ti. The noise cost is also not insignificant, as GTX 1080FE has to ramp up to 52.2dB(A), a 4.6dB(A) increase in noise. Meanwhile FurMark essentially confirms these findings, with a smaller power increase but a similar increase in noise.

As for the GTX 1070FE, neither the increase in power consumption nor noise is quite as high as GTX 1080FE, though the performance uplift is also a bit smaller. The power penalty is just 21W at the wall for Crysis 3 and 38W for FurMark. This translates to a 2-3dB(A) increase in noise, topping out at 50.0dB for FurMark.

Power, Temperature, & Noise Final Words
POST A COMMENT

200 Comments

View All Comments

  • TestKing123 - Wednesday, July 20, 2016 - link

    Then you're woefully behind the times since other sites can do this better. If you're not able to re-run a benchmark for a game with a pretty significant patch like Tomb Raider, or a high profile game like Doom with a significant performance patch like Vulcan that's been out for over a week, then you're workflow is flawed and this site won't stand a chance against the other crop. I'm pretty sure you're seeing this already if you have any sort of metrics tracking in place. Reply
  • TheinsanegamerN - Wednesday, July 20, 2016 - link

    So question, if you started this article on may 14th, was their no time in the over 2 months to add one game to that benchmark list? Reply
  • nathanddrews - Wednesday, July 20, 2016 - link

    Seems like an official addendum is necessary at some point. Doom on Vulkan is amazing. Dota 2 on Vulkan is great, too (and would be useful in reviews of low end to mainstream GPUs especially). Talos... not so much. Reply
  • Eden-K121D - Thursday, July 21, 2016 - link

    Talos Principle was a proof of concept Reply
  • ajlueke - Friday, July 22, 2016 - link

    http://www.pcgamer.com/doom-benchmarks-return-vulk...

    Addendum complete.
    Reply
  • mczak - Wednesday, July 20, 2016 - link

    The table with the native FP throughput rates isn't correct on page 5. Either it's in terms of flops, then gp104 fp16 would be 1:64. Or it's in terms of hw instruction throughput - then gp100 would be 1:1. (Interestingly, the sandra numbers for half-float are indeed 1:128 - suggesting it didn't make any use of fp16 packing at all.) Reply
  • Ryan Smith - Wednesday, July 20, 2016 - link

    Ahh, right you are. I was going for the FLOPs rate, but wrote down the wrong value. Thanks!

    As for the Sandra numbers, they're not super precise. But it's an obvious indication of what's going on under the hood. When the same CUDA 7.5 code path gives you wildly different results on Pascal, then you know something has changed...
    Reply
  • BurntMyBacon - Thursday, July 21, 2016 - link

    Did nVidia somehow limit the ability to promote FP16 operations to FP32? If not, I don't see the point in creating such a slow performing FP16 mode in the first place. Why waste die space when an intelligent designer can just promote the commands to get normal speeds out of the chip anyways? Sure you miss out on speed doubling through packing, but that is still much better than the 1/128 (1/64) rate you get using the provided FP16 mode. Reply
  • Scali - Thursday, July 21, 2016 - link

    I think they can just do that in the shader compiler. Any FP16 operation gets replaced by an FP32 one.
    Only reading from buffers and writing to buffers with FP16 content should remain FP16. Then again, if their driver is smart enough, it can even promote all buffers to FP32 as well (as long as the GPU is the only one accessing the data, the actual representation doesn't matter. Only when the CPU also accesses the data, does it actually need to be FP16).
    Reply
  • owan - Wednesday, July 20, 2016 - link

    Only 2 months late and published the day after a different major GPU release. What happened to this place? Reply

Log in

Don't have an account? Sign up now