Benchmarking Methodology for AI Inference Serving Network Fabrics

This document defines benchmarking terminology, methodologies, and Key Performance Indicators (KPIs) for evaluating Ethernet-based AI inference serving network fabrics. As Large Language Model (LLM) inference deployments scale to disaggregated prefill/decode architectures spanning hundreds or thousands of accelerators (GPUs/ XPUs), the interconnect fabric becomes the critical bottleneck determining Time to First Token (TTFT), Inter-Token Latency (ITL), and aggregate throughput in tokens per second (TPS). This document establishes vendor-independent, reproducible test procedures for benchmarking fabric-level performance under realistic AI inference workloads. Coverage includes RDMA-based KV cache transfer between disaggregated prefill and decode workers, Mixture-of-Experts (MoE) expert parallelism AllToAll communication, request routing and load balancing for inference serving, congestion management under bursty inference traffic patterns, and scale/soak testing. The methodology enables direct, equivalent comparison across implementations, NIC transport stacks (RoCEv2, UET), and fabric architectures. This document is a companion to [TRAINING-BENCH], which addresses training workloads.