Is Tesla K40 clashing with K20 in a multi-GPU system?

I have a multi-GPU system with 4 GPU. Two Tesla K40 and two K20 sitting on the same PCI bus…

When I launch one MPI rank per GPU, I can use one K40 or two K40 or K40 combined with K20… But I cannot use K20 on its own.

Any idea whats happening?