When in doubt, at Microsoft, we would always say the answer is E) All of the above. This machine needs to be "every use" sitting in a lab of a scientific college and also used remotely in the classroom by professors to demonstrate models or setup of models or programming.
A "surevey" of personal super computing may include apps that almost never touch the GPU. OTOH, the GPU may get blistering hot animating Toy Story in real time for which, I think, it needs all lanes. Why older hardware? It has more versatility to teach parallel processing and we just happen to have them on hand. It puts some limits (no SLI, not massively powerful) yet incredibly diverse usage.
It is amazing to me how deep the group is willing to go to answer questions. I will be much more specific when I post. About to post another question that I can find no answers online and the manufacturer refuses to answer as they only directly support 2 specific processors. lol. Hack my own BIOS if I have to.
In answer to a question above, there is no budget and yet the budget is unlimited. These may go in an enclosure that hooks to the PCIe bus directly or via Thunderbolt to a blade system. It may be the same and hook up to a simple ROG GR8 II which is a pretty amazing machine for the $. They may be added to the botherboard of a Dell T7920 desktop with dual scalable processors or a simple 2-way E5-26xx v1/2/3/4 motherboard. The consensus of need is to mainly teach CUDA and do rendering. If we can do more with powerful CPU's, AWESOME.
In the mean time, give this Doc a quick overview?
1) Mining can be done on 1 lane why? Because it runs in a loop and only reports back when it has data?
2) What else runs well on 1x?
3) What applications are the most taxing on BOTH the CPU and GPU?
4) What 3d applications are not able to utilize a graphics card much at all?
5) What applications mainly rely on the maximum number of GPU cores?
6) How much does actively running a display slow down a GPU (should we consider PCioE given the use case stated above and have a portable thin client with multiple monitors that gets wheeled from lecture hall to lecture hall, demos, auditorium, dorm rooms, etc. as scheduled?)