Actually Tesla is and has been based on standard Nvidia GPU technology.
Further, Nvidia was consulted during the design of CAPI and there is nothing preventing Tesla from being implemented via that interconnect, at many times the bandwidth you are quoting.
Nvidia, with its NVLink technology, and IBM with CAPI are creating coherent interconnects with 2.5 to 10 times the bandwidth that Intel hardware can provide.
As to applicability, I'm not sure that either idea has much utility for a desktop user.
But it is much faster than merely plugging Tesla boards into PCI-e slots on an X64 system.
This entire thread, and the idea of desktop "supercomputers" borders on ludicrous as there aren't any real applications for this technology.
But, if and when somebody uncovers these applications, the hardware I've mentioned will be able to run circles around the X64 hardware you've mentioned.
And its on hardware that is much closer to our current platforms than PCs.
That seems to frustrate you, but these are just a statistically provable facts.