schmidtbag:

The only reason I can think why a GPU would have x16 slots nowadays is for backward compatibility. So take a GTX 1070 for example, where you should be able to run that on PCIe 3.0 on x8 lanes and hardly ever see any performance loss. Run it on x4 lanes and you most certainly will. Remember that bandwidth is roughly doubled after every PCIe generation. So, if you stick that GPU in a PCIe 2.0 x16 slot, it should pretty much run at full speed. You're probably saturating almost all of the available bandwidth, but it will run at full speed. Stick it in an x8 slot and you will see a performance loss, since that's equivalent to a 3.0 @ x4.

Compute workloads can also benefit from the extra bandwidth in certain situations.

yasamoka:

Compute workloads can also benefit from the extra bandwidth in certain situations.

Most of the time, it's the exact opposite: heavy compute loads tend to need fewer lanes, not more. So yes, I take your point that certain situations would call for that, but in such situations, you're probably not strapped for 32GB/s of throughput.