EVGA GTX 1070 FTW Compute Benchmarks
Compute can still be a good reason to buy a consumer GPU. They can provide a majority of the performance compared to their professional counterparts just with a few tradeoffs that may be acceptable for some use cases. They can offer a lot of good value that shouldn’t be dismissed for those amateur or budding developers looking for a compute card.
Here we’re testing Caffe, which isn’t quite FP16 yet, though is still a very viable test of compute capability. But just how much better is this going to be in comparison to it predecessors, or even with its rivals. Caffe DNN can easily be run by anyone on a *nix distribution. Simply download the applicable binaries, compile it with either cuDNN or use AMD’s open source OpenCL version. You can then test the benchmark that’s inherent in the Caffe framework.
Next up we’ll test the familiar, and updated, Folding@Home benchmark which is a test of single and double precision work. It’s a good benchmark that tests a specific workload, though it does give a nice ideaa of how it performs.
Monte Carlo Financial Simulation
SiSoft Sandra has a nice financial OpenCL compute test that runs through a Monte Carlo simulation of monetary futures. It is, again, a specific workload but it does give a great idea of how well the GTX 1070 FTW can do with potentially real-world scenarios.
To test we ran the Ashes of the Singularity benchmark at the Extreme preset at 2160P. We measured PCIe power using leads connected to the slot itself, as well as the power connector. These are are the peak numbers that were seen.
Pascal is certainly very competitive in terms of actual power consumption and the GTX 1070 FTW, despite being clocked higher, consumes less power than a GTX 970 and is only beaten out by our sample of the RX 480. Not bad indeed.
Temperature was taken using readings from HWiNFO64 while playing through the Ashes of the Singularity benchmark at Extreme at 2160P. We then monitored peak temperature throughout and these are those peak numbers.
These temperatures are solely of the GPU die itself, and the cooling solution is more than adequate at keeping things acceptable. The higher idle temperatures are due to the fan not being on at all, and it’s well with tolerances. VRM’s are a different matter entirely, though I think the situation has been blown out of proportion. During testing we never saw temperatures over 96-degrees, which is hot but within specifications.
VRM’s, when properly made and sourced, can indeed withstand the heat from normal operations. These are rated well above the temperatures that they are facing, though the board surrounding may not be. And it’s when we push well beyond the specifications of the greater board that we may come into trouble. That is, when you overclock the chip itself by a larger margin. In FTW trim it performs flawlessly without danger of exploding. There is cooling, with the fan blowing air directly through the heatsink and onto them, though it doesn’t appear to be enough and has been deemed an engineering fallacy. Simple heat pads seem to do wonders, though EVGA should indeed have added something else to help control heat on those vital components. It works fine as is, but it just doesn’t overclock without having problems. Because of that, we don’t recommend overclocking your ACX 3.0 wearing EVGA cards unless you have more
For sound we measured at 1 meter away from an open chassis while we played the aforementioned Ashes of the Singularity benchmark. We used a Reed R8050 to test. We also used an iPhone 6 Plus just to see if it was accurate.
Thankfully the zero DB idle, or having the fan turn off completely is a very useful feature. While browsing around on the desktop and the Internet you shouldn’t have any noise from it at all. At load, 33dBa doesn’t sound annoying at all, either. It’s among the quietest cards there are. Of course you can make the fan curve more aggressive, but why would you when it already works great at stock.