So they pitched an i9-10900K against an i7-10750H? A 10C/20T 5.3GHz CPU versus a 6C/12T 5GHz one? A 125W TDP chip pitted against a 45W TDP processor? At 1080p?
And there was a performance difference? Well colour me surprised...
Sarcasm aside, a far more appropriate setup would have been to use the same desktop system, and compared the RTX 3090 in the motherboard against the external unit. There are plenty of LGA1200 motherboards that support Thunderbolt 3.
While certainly not a close comparo between processors, those actual differences are directly relevant to what was tested. In theory I'd love a direct comparison with identical processors, just to see what the bandwidth reduction and latency penalties are with TB3 in different games, but that's not relevant when you actually use a TB3 eGPU.
Because you won't be using a TB3 enclosure for your GPU on any LGA1200 motherboard. Why would you?
Instead, many people will be using a TB3 with an i7-10750H or similar, that's a middle of the road processor option for a TB3 laptop someone would be considering an eGPU for. There are some less powerful 4c8t and more powerful 8c16t systems, so this 6c12t i7-10750H is a reasonable option to test with.
And the results certainly suggest that spending on the highest tier GPUs for an eGPU setup is just throwing money away (current ridiculous pricing aside). Spending the same on an eGPU w/3080 ($350+$800, realistic eventual price?) and a desktop w/3060Ti ($650+$500) will net you more FPS on the desktop and greater headroom for performance upgrades in the future.
Yes the 1 computer advantage is real nice, but IMO the performance loss at the top as tested by KG suggests that money spent here is mostly wasted. IMO the performance loss with my 1080 is not acceptable (GPU money wasted) but the smaller loss seems OK at around the 1060 level.
I'd like to get my 1660 Super (and the 1080) in the TB case again now that I recently acquired a 6c6t TB3 Core i5-8500B Mac Mini (booted to Win 10) to compare to my 6c6t Core i5-8400 PC. 0.1 GHz difference in ACT and I can manually match the RAM timings, so this will be about as close as I can get to an apples-to-apples comparo.