Overclocking, Power Consumption, Noise, and Temps
We also spent some time overlcocking the new GeForce GTX 1070, to see what kind of additional performance we could squeeze out of it. Before we get to our results, though, we need to talk about an update to GPU Boost and some new overclocking related features that debuted with Pascal.
Pascal-based GeForces like this card and the GTX 1080 feature GPU Boost 3.0. Like previous-gen GeForce cards, GPU Boost scales frequencies and voltages, based on available power and temperature limitations, in-line with the GPU's workload at the time. With GPU Boost 2.0, which is available on older Maxwell-based cards, you could overclock using a fixed frequency offset and tweak the GPU voltage curve manually, but that left some performance on the table because it doesn't offer the kind of granular control necessary to massage every last MHz out of a given chip at every voltage point.
With GPU Boost 3.0, however, you can now manually set the frequency offset on a per voltage point basis if you so choose. In addition, a new feature coming to overclocking / monitoring utilities like EVGA's Precision X gives users the ability to run a tool that automatically finds a GPU's peak frequency at each available voltage point. The tool basically runs Furmark and tests for stability, while incrementally upping the GPU frequency per voltage step point. When instability or image anomalies are detected, the tool logs the peak "good" frequency and moves on to the next voltage step. You eventually end up with a customized overclocking profile, optimized for your particular GPU.
Unfortunately, due to an issue which caused V-Sync to be enabled after every re-boot, we couldn't properly test the auto-scan feature, which required a couple of re-boots, so we kept things simple, and used the frequency offset and power / temperature target sliders also available in Precision X. First we cranked up the power target to 120% and then kicked up the temperature target to 91 degrees. Then we bumped up the GPU and memory clocks until the test system was no longer stable.
Ultimately, we saw a peak GPU clock of over 2GHz with a 4090MHz memory clock (effective speed 8.1Gbps). Those are some impressive GPU clock speeds, given the relative immaturity of TSMC's 16nm FinFET process and the 4-phase power circuitry on the 1070. When yields improve on the GP104 (assuming they will) and board partners slap additional power feeds and bigger coolers on GTX 1070 cards, higher clocks may be possible.
The GeForce GTX 1070 consumed about the same amount of power as the Radeon R9 Nano, despite offering better overall performance, and markedly less power than all of the other cards we tested.
We should mention that temperatures on the GeForce GTX 1070 are a non-issue as well. At idle, the card's GPU hovered in the high-30s. And under sustained load, the GPU temp shot up into the high 80'C range, right in-line with the default 83'C target. While overclocked though, it'll usually shoot right up to wherever the temperature target is configured.
In terms of its noise output, the GeForce GTX 1070 is very quiet. When idling and under load, it proved to be among the quietest cards of the bunch. We should note, however, that much of the additional noise under load in this graph can be attributed to our system's PSU and CPU cooler. All of these graphics cards are relatively quiet under load, especially if you test inside a chassis like we do.