To this end, that also means that declaring hard victors is sort of rough – all those statements are made based on our samples, but it’s always possible that one card might perform +1% in our sample, but -1% in another sample. Among other reasons, this is why gaming performance gets a back seat once we’ve determined a baseline for AIB partner performance of a new GPU.

What matters more is the cooling solution, the PCB quality, and the noise levels. We are therefore checking for things like component quality (already done separately, but revisited below), thermals, and noise output. Gaming, again, gets a back seat. Overclocking is also inevitably limited by Pascal itself, so even that is relatively unexcited for same-GPU video card comparisons.

EVGA GTX 1080 Ti FTW3 PCB, VRM, & Tear-Down

We’ve already posted videos showing a tear-down of the EVGA 1080 Ti FTW3 (found here) and analyzing in great depth the VRM & PCB of the FTW3.

Recapping the most important bits: The GTX 1080 Ti FTW3 uses an excellently engineered PCB and power delivery solution, accompanied by an equally over-engineered cooling solution. The Vcore VRM is comprised of Alpha-Omega Semiconductor E6930 Dual-N MOSFETs, which package high-side, low-side, and diode components into the same package. There are twenty total FET packages on the board, ten driver ICs, and five doublers, combined with an NCP81274 voltage controller. The voltage controller is an 8-phase controller capable of 1.2MHz switching frequency. EVGA is using NCP81162 doublers to load balance between the two phases, largely eliminating the risk of out-of-balance current between two phases (as they are balanced by EVGA’s doublers). This also reduces strain on the 12V rail, which should make the power supply happy.

EVGA’s drivers are also from NCP, using NCP81158 drivers, though they can’t quite keep up with the rise and fall times of the dual-N FET rise and fall times (and can’t support higher gate-drive voltages than 5v).

Learn more about all of this in our PCB breakdown video:

As for the rest of the card, it’s largely comprised of overcompensating cooling solutions across all devices – yes, even fan controllers are connected via thermal pad – and three fans to individually cool partitions of the card. The heatsink is split into two primary halves – GPU+MEM and PWR – and uses a mix of EVGA’s new fin designs to improve cooling efficiency. Although the fin density and placement aren’t all that unique, the tail of the fins can be: EVGA switches between straight fins (with no contact to componentry), L-shaped fins (with contact to thermal pads, but permitting airflow), and closed fins (for full contact, but no airflow). The type of fin used depends on which component that part of the cooler covers; VRM components tend to contact the finstack via L-shaped fins, whereas less critical inductors are under straight fins.

Several MCUs are on the card for individual and asynchronous fan control, slaving to one of EVGA’s somewhat aggressive fan profiles (more on that momentarily), and the MCUs report to a functional RGB LED GPM meter on the card. GPM means “GPU,” “Power,” and “Memory.” This card is laid-out in a fashion that it’s more accurate to say GMP. Graphics and memory fans are on the left, power is on the right.

As for thermistors used, we previously detailed in great depth that EVGA is using negative type thermistors (NTCs) in 9 locations on the board. What’s important here is that the location of those thermistors is not the same between EVGA ICX devices, so PWR1 does not necessarily equal PWR1 between two cards. The SC2 card, for instance, uses a reference nVidia FE PCB with EVGA’s NTC thermistors positioned all along the front of the board, while the FTW3 PCB positions a few on the back-side. PWR4 is the biggest difference, which we’ll talk about in the thermal section.

You cannot compare EVGA ICX thermistor readings between cards without scrutiny.

The two can be compared, but only insofar as GPU and MEM readings (and even then, only if EVGA never changes which memory modules receive which reporting tag).

GPU Testing Methodology

For our benchmarks today, we’re using a fully rebuilt GPU test bench for 2017. This is our first full set of GPUs for the year, giving us an opportunity to move to an i7-7700K platform that’s clocked higher than our old GPU test bed. For all the excitement that comes with a new GPU test bench and a clean slate to work with, we also lose some information: Our old GPU tests are completely incomparable to these results due to a new set of numbers, completely new testing methodology, new game settings, and new games being tested with. DOOM, for instance, now has a new test methodology behind it. We’ve moved to Ultra graphics settings with 0xAA and async enabled, also dropping OpenGL entirely in favor of Vulkan + more Dx12 tests.

We’ve also automated a significant portion of our testing at this point, reducing manual workload in favor of greater focus on analytics.

Driver version 378.78 (press-ready drivers for 1080 Ti, provided by nVidia) was used for all nVidia devices. Version 17.10.1030-B8 was used for AMD (press drivers).

A separate bench is used for game performance and for thermal performance.

Thermal Test Bench

Our test methodology for the is largely parallel to our EVGA VRM final torture test that we published late last year. We use logging software to monitor the NTCs on EVGA’s ICX card, with our own calibrated thermocouples mounted to power components for non-ICX monitoring. Our thermocouples use an adhesive pad that is 1/100th of an inch thick, and does not interfere in any meaningful way with thermal transfer. The pad is a combination of polyimide and polymethylphenylsiloxane, and the thermocouple is a K-type hooked up to a logging meter. Calibration offsets are applied as necessary, with the exact same thermocouples used in the same spots for each test.

Torture testing used Kombustor's 'Furry Donut' testing, 3DMark, and a few games (to determine auto fan speeds under 'real' usage conditions, used later for noise level testing).

Our tests apply self-adhesive, 1/100th-inch thick (read: laser thin, does not cause "air gaps") K-type thermocouples directly to the rear-side of the PCB and to hotspot MOSFETs numbers 2 and 7 when counting from the bottom of the PCB. The thermocouples used are flat and are self-adhesive (from Omega), as recommended by thermal engineers in the industry -- including Bobby Kinstle of Corsair, whom we previously interviewed.

K-type thermocouples have a known range of approximately 2.2C. We calibrated our thermocouples by providing them an "ice bath," then providing them a boiling water bath. This provided us the information required to understand and adjust results appropriately.

Because we have concerns pertaining to thermal conductivity and impact of the thermocouple pad in its placement area, we selected the pads discussed above for uninterrupted performance of the cooler by the test equipment. Electrical conductivity is also a concern, as you don't want bare wire to cause an electrical short on the PCB. Fortunately, these thermocouples are not electrically conductive along the wire or placement pad, with the wire using a PTFE coating with a 30 AWG (~0.0100"⌀). The thermocouples are 914mm long and connect into our dual logging thermocouple readers, which then take second by second measurements of temperature. We also log ambient, and apply an ambient modifier where necessary to adjust test passes so that they are fair.

The response time of our thermocouples is 0.15s, with an accompanying resolution of 0.1C. The laminates arae fiberglass-reinforced polymer layers, with junction insulation comprised of polyimide and fiberglass. The thermocouples are rated for just under 200C, which is enough for any VRM testing (and if we go over that, something will probably blow, anyway).

To avoid EMI, we mostly guess-and-check placement of the thermocouples. EMI is caused by power plane PCBs and inductors. We were able to avoid electromagnetic interference by routing the thermocouple wiring right, toward the less populated half of the board, and then down. The cables exit the board near the PCI-e slot and avoid crossing inductors. This resulted in no observable/measurable EMI with regard to temperature readings.

We decided to deploy AIDA64 and GPU-Z to measure direct temperatures of the GPU and the CPU (becomes relevant during torture testing, when we dump the CPU radiator's heat straight into the VRM fan). In addition to this, logging of fan speeds, VID, vCore, and other aspects of power management were logged. We then use EVGA's custom Precision build to log the thermistor readings second by second, matched against and validated between our own thermocouples.

The primary test platform is detailed below:

Note also that we swap test benches for the GPU thermal testing, using instead our "red" bench with three case fans -- only one is connected (directed at CPU area) -- and an elevated standoff for the 120mm fat radiator cooler from Asetek (for the CPU) with Gentle Typhoon fan at max RPM. This is elevated out of airflow pathways for the GPU, and is irrelevant to testing -- but we're detailing it for our own notes in the future.

Game Bench 

BIOS settings include C-states completely disabled with the CPU locked to 4.5GHz at 1.32 vCore. Memory is at XMP1.

We communicated with both AMD and nVidia about the new titles on the bench, and gave each company the opportunity to ‘vote’ for a title they’d like to see us add. We figure this will help even out some of the game biases that exist. AMD doesn’t make a big showing today, but will soon. We are testing:

• Ghost Recon: Wildlands (built-in bench, Very High; recommended by nVidia)
• Sniper Elite 4 (High, Async, Dx12; recommended by AMD)
• For Honor (Extreme, manual bench as built-in is unrealistically abusive)
• Ashes of the Singularity (GPU-focused, High, Dx12)
• DOOM (Vulkan, Ultra, 0xAA, Async)

Synthetics:

• 3DMark FireStrike
• 3DMark FireStrike Extreme
• 3DMark FireStrike Ultra
• 3DMark TimeSpy

For measurement tools, we’re using PresentMon for Dx12/Vulkan titles and FRAPS for Dx11 titles. OnPresent is the preferred output for us, which is then fed through our own script to calculate 1% low and 0.1% low metrics (defined here).

Power testing is taken at the wall. One case fan is connected, both SSDs, and the system is otherwise left in the "Game Bench" configuration.