Ultra Ethernet Link Layer Retry (LLR) and Credit-based Flow Control (CBFC) Test Solution
技术资料
As AI clusters scale, vendors are replacing proprietary chip-to-chip interconnects with open Ethernet inside the rack — the ‘scale-up’ domain linking GPUs at extreme bandwidth and sub-microsecond latency. With short reach and only one or two hops, the link itself must carry reliability: a single-bit error or buffer overrun can stall a synchronized training step across thousands of GPUs. Link layer reliability is becoming a gating requirement for any silicon, NIC, or switch targeting scale-up AI infrastructure. Ultra Ethernet 1.0 adds two link-layer mechanisms built for AI and HPC fabrics. Link Layer Retry (LLR) detects frame errors at the link and retransmits locally in microseconds, rather than relying on higher-layer recovery. Credit-Based Flow Control (CBFC) prevents buffer overflow and loss by having receivers advertise buffer credits per virtual channel (VC), so senders transmit only when credits allow. Together they deliver near-lossless, low-latency behavior that keeps GPU usage high. Keysight's hardware-based LLR / CBFC test solution validates implementations against the UE 1.0 specification — from link bring-up and steady-state traffic to targeted fault and congestion injection — at full line rate up to 1600GE. Spec-mapped statistics, enhanced preamble capture, and PCS capture help pinpoint and troubleshoot link-layer issues quicker.