...PCI-E 3.0 reaches about 256 Gbps, while PCI-E 512 Gbps. Any future card you might want to add to an iMP would be more powerful than a Vega, which means very likely that it will have a high bandwidth workflow, exploiting fully PCI-E.....I seriously doubt that your money would be well spent with an ...1180 Nvidia or an AMD Vega 2 card in a TB3 enclosure. At least if you're looking to improve performance on an iMP.
There is already good data that increasing PCI-E bandwidth has less impact on real-world performance than you'd think.
Adding PCI-E lanes has almost no benefit: http://www.gamersnexus.net/guides/2488-pci-e-3-x8-vs-x16-performance-impact-on-gpus
PCI-E 3.0 has limited benefit over 2.0: https://www.pugetsystems.com/labs/a...E-Speed-on-Gaming-Performance-518/#Conclusion
PCIE- 3.0 has limited benefit over 2.0: http://www.hardwaresecrets.com/pci-express-3-0-vs-2-0-gaming-performance-gain/5/
Thunderbolt 3 has varying but often modest performance loss vs PCI-E. Note this is still much faster than the original GPU. The point is whether the improvement from baseline is beneficial for the task at hand, not whether it's exactly equal to PCI-E.: https://egpu.io/forums/mac-setup/pcie-slot-dgpu-vs-thunderbolt-3-egpu-internal-display-test/
At the WWDC discussion on eGPU support, Apple stated the performance situation is highly variable and there's no single accurate depiction. Some apps are designed to do lots of work on the GPU with fewer API calls, in which case there is minimal TB3 perf. penalty. Other apps do lots of little calls to the GPU which would have more overhead on TB3. Furthermore the efficiency and utility of eGPU support will be constantly changing in the future as developers use the new macOS features that support this.
Due to these varying parameters, it is unknown whether future eGPU cards added to an iMP would not be cost effective due to lack of TB3 bandwidth. It depends on the exact workflow and how that future app is written to use the GPU. Today nobody knows that with any confidence.