not sure what you heard, but cudaminer is still using quite a lot of PCI express bandwidth.

At minimum, all the generated hashes (32 bytes per hash) are sent back to the CPU for
evaluation (-H 2). When using -H 1, 128 bytes per hash are sent to the GPU and back
before and after scrypt

-H 1 uses 8 times more bandwidth than -H 2, but even -H 2 still  uses a significant amount.

The only exception is Maxcoin where hashes are already evaluated on GPU.

Now that Maxwell has gathered significant interest and people are trying to build 5 and
6 GPU systems with risers, it would be a good time to fix this.

Christian