draft-calabria-bmwg-ai-fabric-terminology
Benchmarking Terminology for AI Network Fabrics
This document defines benchmarking terminology for evaluating
Ethernet-based network fabrics used in distributed Artificial
Intelligence (AI) training and inference workloads. It provides a
unified vocabulary consolidating and extending terms from RFC 1242,
RFC 8238, and the companion AI fabric methodology documents,
establishing precise, vendor-neutral definitions for collective
communication primitives, RDMA transport mechanisms (RoCEv2 and Ultra
Ethernet Transport), congestion control behaviors, AI-specific Key
Performance Indicators (KPIs), and fabric topology concepts.
This document is a companion to
[I-D.calabria-bmwg-ai-fabric-training-bench] and
[I-D.calabria-bmwg-ai-fabric-inference-bench]. Those documents
SHOULD NOT be applied without first consulting the terminology
defined herein. Where definitions herein overlap with RFC 1242 or
RFC 8238, the AI fabric context definition in this document takes
precedence.