GPU-initiated, fine-grained RDMA communication, e.g., NVIDIA IBGDA, lets GPU threads directly issue small RDMA requests to the RNIC. IBGDA’s fine-granularity benefits MoE communication pattern, as exemplified by current art, DeepEP from DeepSeek, employing it; but IBGDA requires NVIDIA’s GPUs and NICs, worsening the vendor lock-in.

UCCL-EP presents a GPU-initiated, but CPU-proxied communication to enable such fine-grained RDMA communication via CPU w/o needing specific GPU-NIC integration. GPU threads inject RDMA requests to UCCL-EP established GPU-CPU channel, and CPU threads issue the corresponding GPUDirect RDMA requests via Linux libibverbs. Similar to UCCL paper, UCCL-EP convinces that CPU driving the control path, not the data path, of RDMA can preserve good performance and more flexibility at the same time.