Partner Cards And Efficiency Testing
A number of custom GeForce GTX 1070 board and cooling designs are now available (at least to back-order; supply of Pascal-based hardware remains sparse overall). So, we’re pitting a few different cards against each other. First, we have Nvidia's GeForce GTX 1070 and 1080 Founders Edition. Then there's MSI's GeForce GTX 1070 and 1080 Gaming X 8G. The last card we're using to compare is MSI's GeForce GTX 980 Ti Lightning, powered by the Maxwell architecture.
Relationship Between Clock Frequency, Power Consumption & Performance
Although we're running a great many tests, the work is worthwhile in order to determine each card's efficiency. This process involves measuring power consumption in steps across a wide range of clock rates, then comparing that information to gaming performance. By the end, we'll know whether overclocking or dialing back power are worth pursuing. We're also comparing Maxwell and Pascal to each other after correcting for clock rate, which should shed some light on the performance of each CUDA core and how they interact with the GPU's other subsystems.
It goes without saying that the next logical step after a detailed analysis of power consumption is a deeper dive into temperature and noise. But before detailing our setup and methodology, let's take a quick look at the technical specifications of all five graphics cards:
Gaming X 8G
Test Methodology & Benchmark Selection
In order to maintain as close to real-world conditions as possible, we ran all of our tests in a closed Nanoxia Deep Silence 5 chassis. Its stock fans were set to spin as slowly as possible in the front and at their medium setting in the back.
Power consumption is measured according to the processes outlined in The Math Behind GPU Power Consumption And PSUs. Positioning the current probes proved a bit of a challenge this time, but we made it work.
|Test Method||Contact-free DC Measurement at PCIe Slot (Using a Riser Card) Contact-free DC Measurement at External Auxiliary Power Supply Cable Direct Voltage Measurement at Power Supply|
|Test Equipment||2 x Rohde & Schwarz HMO 3054, 500 MHz Digital Multi-Channel Oscilloscope with Storage Function 4 x Rohde & Schwarz HZO50 Current Probe (1 mA - 30 A, 100 kHz, DC) 4 x Rohde & Schwarz HZ355 (10:1 Probes, 500 MHz) 1 x Rohde & Schwarz HMC 8012 Digital Multimeter with Storage Function 1 x Optris PI640 80Hz Infrared Camera + PI Connect|
The next challenge was choosing a benchmark. Because we'd need to run so many data points, we could only pick one test to use. Even then, collecting all of the information we needed took almost six full work days.
We tried to be as representative as possible. To this end, we performed a number of preliminary tests and, based on their results, ended up choosing Metro: Last Light at Ultra HD resolution. This allowed us to run the graphics cards at full load without running into CPU or memory bottlenecks, meaning we're looking at the worst-case scenario for graphics card power consumption without causing GPU Boost to pull clock rates lower.
As a foundation for our platform, we used a Core i7-6700K processor running at 4.5 GHz under an all-in-one liquid cooler. This way, we made sure that the rest of the system wasn’t going to negatively influence our test results.
|Happyware Crossover Workstation|
|Test System||Core i7-6700K @ 4.5 GHz2x 8 GB DDR4-3400Asus P10 WS2x Samsung SM863 (3D V-NAND)Seagate Constellation Server HDD Windows 10 Enterprise (TH2, All Updates)|
MORE: Best Graphics Cards
MORE: All Graphics Content
If it were me, I would have removed a low to middle clock rate instead of the very lowest to get both the low idle power consumption and the OC speed.
Also, about overclocking: I think reviews of all these new generation nVidia and AMD cards should include average clock that cards operated when doing all game benchmarks. Official boost clock numbers are a bit useless, because AMD cards run games at below boost clocks, and average for nVidia GTX1070 is above boost clocks. Having just official boost clock numbers make it difficult to evaluate overclocking potential and make real gains look much bigger or smaller than expected.