InfiniBand

Overview

InfiniBand is a high-performance networking interconnect standard originally designed for HPC environments. Its volume one specification runs to over 2,000 pages; the InfiniBand transport specification was standardized in 2000 and served as the basis for RoCEv2 (RDMA over Converged Ethernet). InfiniBand's design — optimized for a transistor-expensive era — carries go-back-N error recovery, no multipathing, and no integrated security, all of which have become structural limitations as AI clusters have scaled to millions of GPUs. The Ultra Ethernet Consortium's specification 1.0, at just over 500 pages, achieves comparable coverage by building on top of existing Ethernet and IP ecosystems rather than reinventing the full stack.

Sign in to read the full article.

Sign In