2017 Intel NUC7i5BNH [7th,2C,U] + GTX 1080 Ti @ 32Gbps-TB3 (Sonnet Breakaway 550) + Win10 [hkoster1]
So here are my build details for the NUC7i7BNH + GTX 1080Ti.
Intel NUC7i7BHN PC kit with Core i7 (7th gen) CPU and, among others, Thunderbolt 3 (max 40Gbps) and HDMI ports, plus onboard Iris Plus Graphics 650. I have completed this NUC with 16GB DDR4 RAM (for now), a 500GB Samsung 960 PRO NVME SSD in the m2-slot (for Windows 10 and applications), and a 500GB Samsung 860 EVO SSD in the drive bay with SATA-3 slot (for bulk storage).
My 4K screen is a Dell P2415Q, initially connected to the NUC via HDMI cable, and later to the eGPU via mDP-to-DP cable.
My eGPU enclosure is the Sonnet eGFX 550W, equipped with an MSI GeForce GTX 1080Ti GamingX 11GB DDR5 video card. The eGPU came with a 50cm TB3 cable, which is OK for the stationary NUC. For connection to a laptop (like my MacBook Pro) I also have the 80cm cable sold by Apple.
The Dell P2415Q screen is placed (in its lowest position) on top of the Sonnet box. I use the NUC/eGPU setup mainly for gaming, so it doesn't matter that it is at arm's length distance. That leaves room in front of it for working on my MacBook Pro. Note the Bluetooth Logitech K810 Windows keyboard and M557 mouse that connect to the NUC when in use.
The HWINFO panel shows details of the NUC itself.
I ran Intel's Thunderbolt 3 Firmware Update Tool on the NUC, but aborted it when the installed version 33 appeared to be up-to-date. Installing the latest NVidia GeForce driver package presented a problem (to this Windows novice) in that the procedure stopped halfway through because it could not find suitable NVidia hardware - on the NUC, that is. This remains the case even when the eGPU is up and running.
Luckily, Windows 10 recognizes the GTX 1080Ti in the eGPU when switched on and attached via TB3 cable. Device Manager offered to look on the internet for a suitable driver, found one (possibly not the latest version) and installed it automatically. I'll continue to look for a way to install the latest driver package.
At this stage I had two video connections from the NUC to the Dell screen: one directly to the HDMI port; and one via the eGPU to the mDP port. I could switch between these inputs on the Dell screen, but only the HDMI connection was of any use. It appeared in Settings > Display that Windows 10 recognized the presence of two screens, with screen 2 being the eGPU-accellerated screen.
Making this screen the default solved the problem (after a reboot), after which I could then disconnect the HDMI cable from the NUC. There is still a small inactive screen 1 shown in Settings > Display, next to the big active screen 2. Note also that the Dell P2415Q is now once-removed from the NUC, and Windows only recognizes it as a generic PnP screen. I found the scaled 1920x1080 resolution to work best for my purposes.
There was a bit of trial-and-error involved (especially error since I'm not really a Windows person), but I'm now happy with the result.
Before I could run Unigine Heaven 4.0 benchmarks I had to solve another puzzle: where to find the missing MSVCP100.dll without which Heaven will not run. A long story short: install the 2010 version of the Visual C++ redistributable for x86 (even though my NUC runs 64-bit Windows 10). Various more recent versions just will not do...
The Unigine Superposition and Heaven 4.0 benchmarks, as well as the AIDA64 benchmarks speak for themselves. I'm using this NUC/eGPU setup for running N3V Games Trainz (T:ANE) sessions in large routes that stress any CPU/GPU combination, and I'm happy with its performance. The Kaby Lake i7 Core CPU and the NVME SSD are running hot, but not to the point that putting a hand on the top of the NUC is uncomfortable.
This may be an incredibly stupid question but the CPU only has 4 PCIe lanes. I don't understand how this logically works with m.2 and the thunderbolt 3 controller.
So... does using an NVMe drive affect the bandwidth via the thunderbolt 3 bolt? Surely it would? If so, is the impact negligble because I'd assume the difference would be similar to tb2-tb3 which seems to be <10%.