TB3 RAID will outperform both of these.
How?
If I understand correctly, the 8950HK (latest MacBook Pro), or 8700B (latest Mac mini), or 7700k (latest iMac) only has 16 PCIe lane in total. Even that's PCIe 3.0. The total available bandwidth still less than a single W3690 can provide.
AFAIK, iMac Pro (48 PCIe lane) is the only exception that can really provide more bandwidth (than cMP) for multi GPU setup.
However, the block diagram suggest that only 8 lane (in total) for all four TB ports. Therefore, the combined bandwidth for all four TB3 ports just equals to a single PCIe 2.0 x16 slot.
Of course, if you put the iMac Pro's onboard Vega's PCIe 3.0 x16 bandwidth into the calculation, then the combined bandwidth for all GPU will become PCIe 3.0 x24, that's definitely more than the max possible that a single cMP can achieve (PCIe 2.0 x36).
However, it's very clear that the onboard Vega is downclocked to fit inside the iMac Pro's thermal / power envelope. No matter give it how much bandwidth. It still can't perform better than the Vega on cMP. Most of the bandwidth at there is simply wasted.
So, if we look at the real possible GPU performance comparison.
A Vega FE in cMP PCIe slot 1
can do better than the Vega on iMac Pro. Luxmark proved this.
The PCIe slot 2 can provide same bandwidth as all four TB3 (combined) on the iMac Pro. So, if we assume the GPU performance is bandwidth limiting, then they should perform the same in this area.
However, the cMP still has a slot 3 and 4 to accommodate more GPU. iMac Pro? Already run out of expansion bandwidth (for GPU).
If the workflow is really GPU limiting, but not CPU limiting. I am afraid that none of the newer Mac can out perform a 5,1.