NVIDIA links data centers into a unified AI supercomputer with Spectrum-XGS Ethernet

Data center networking is central to distributed computing and to future Artificial Intelligence workloads that may span millions of GPUs. NVIDIA introduced Spectrum-XGS Ethernet as an extension of its Spectrum-X networking platform designed to link multiple, geographically separated data centers into a unified, giga-scale Artificial Intelligence super-factory. The company said Spectrum-XGS removes the capacity limits of single facilities by adding distance-aware networking, which aims to provide predictable, low-latency performance across campuses, cities, and continents.

The changes are delivered primarily through software and firmware updates to existing Spectrum-X switches and ConnectX SuperNICs rather than through new silicon. Spectrum-XGS includes auto-adjusted congestion control tuned for long-haul links, precise latency management to reduce jitter, and comprehensive end-to-end telemetry. That telemetry is intended to allow operators to visualize and control network traffic across multiple sites, giving visibility into cross-facility flows and making behavior across long distances more predictable for distributed workloads.

NVIDIA reported measurable performance improvements from the updates, saying Spectrum-XGS nearly doubles NCCL throughput for multi-GPU, multi-node training jobs and large-scale experiments. Those gains are presented as efficiency improvements for distributed Artificial Intelligence training and inference. NVIDIA positioned the technology as a new axis of growth for infrastructure, following scale-up inside servers and scale-out inside data centers with a new scale-across approach that connects facilities into unified compute fabrics as demand for massive distributed compute grows.