It depends what your requirements are but I've successfully used a load of http://amfeltec.com/products/flexible-x1-pci-express-3-way-splitter/ to connect 3 PCIe graphics cards (with x16 to x1 adapters) to a mini-itx MB with a single x16 port which worked OK (for use with opencl with minimal communication going on).
"for use with opencl with minimal communication going on" <- sure x1 is fine for this.
But for science, why not find out how fast she can go!
There is now an additional product from the same company:
http://amfeltec.com/products/flexible-x4-pci-express-4-way-splitter-gpu-oriented/ Which uses an x4 host card branching out to a total of 4 physical x16 slot adapters (included). HOWEVER these seem to be attached by the same electrically x1 flat flex ribbon cables just like the other product. I was unable to find out which chip they were using; pic is not high res enough and its not itemized in the datasheet, but it is for sure a PLX chip.
![](http://amfeltec.com/wp-content/uploads/2013/11/x4-pcie-splitter4-host-card.jpg)
(Large ,has a heatsink shown elsewhere, and datasheet/specs-page seems to indicate that each of the 4 are all addressable simultaneously and without a bottleneck).
It would be interesting to know how much these solutions cost, and if indeed it would be possible to design a better one, either with an x16 host card branching into 4 x4's, or the original intention of x16 into x8+x8. Either using a faster more expensive PLX chip to copy this solution, or with the quad ASMedia 1480 chips (how i found this thread). I have confidence in the BIOS bifurcation thing being possible.
I see no reason for this hardware build to be impossible, just a matter of why.
-Abei Villafane