RDMA · Remote Direct Memory Access
Network transfers that move data directly between machines' memory, bypassing the CPU for low latency.
Current numbers
>5xCXL KV-pool speedup vs SSD-based caching; ~3.8x vs 200G RDMA; up to 87% less GPU memory used
~1–2.5 usend-to-end RDMA fabric latency (NIC-to-NIC through one switch hop): InfiniBand ~1–2 µs, tuned RoCEv2 ~1.5–2.5 µs — microseconds, vs milliseconds for WAN
June 2025Ultra Ethernet Consortium UEC Specification 1.0 released (UET transport, packet spray, UCCM, packet trimming, native RDMA)