Not quite sure why my comment got downvoted so heavily. My point is a commodity GPU is attached to 4 physical ports, each capable of 80 Gbit/sec for a total ba

Not quite sure why my comment got downvoted so heavily. My point is a commodity GPU is attached to 4 physical ports, each capable of 80 Gbit/sec for a total bandwidth of 320 Gbit/sec with DP 2.0. The GP's comment on using the GPU for offloading is interesting. There's a physical link connected directly to the GPU without having to go over the PCI bus or through the CPU. Bi-directional also doesn't matter. Many applications are uni-directional or async. I could easily see a situation where one DP link is used for upstream and 3 DP links are used for downstream, or even all DP links for one direction and a traditional interface used for upstream.