Clinic GPU RDMA must receive lossless treatment from UCS into the fabric. What UCS QoS choice is required?

Clinic GPU RDMA that must receive lossless treatment from UCS into the fabric requires assigning RDMA traffic to the UCS system class designated for lossless treatment. System class selection drives how adapter traffic is queued and marked toward the PFC/lossless fabric design used by municipal clinic inference and training. Domain 3.2.e requires that assignment for AI RDMA rather than enabling link-level pause flow control on fabric uplinks for the RDMA CoS value, which pauses the entire link instead of providing per-priority lossless treatment. Assigning RDMA traffic to the Fibre Channel lossless system class used for FCoE does not provide the proper lossless queuing and marking for GPU RDMA on the path into the fabric. Enabling link-level pause flow control on fabric uplinks for the RDMA CoS value provides no per-priority lossless classification for RDMA and pauses all traffic on the link under congestion, which disrupts the fabric design. Mapping RDMA traffic to a priority CoS value with jumbo frames on the vNIC template improves payload size but leaves training traffic without lossless queuing and PFC marking, so packets can still drop under congestion, which breaks the lossless requirement. Exam caveat: the lossless system class must still align with fabric PFC priority and QoS maps; picking the class name without fabric pairing is incomplete. Operational check: assign clinic GPU RDMA to the UCS lossless system class, verify markings egress correctly, confirm the adjacent fabric applies PFC to that priority, and run a congestion test showing pause behavior without best-effort drops on the RDMA class.