Particle.news
Download on the App Store

NVIDIA's Spectrum-6 Switch Enters Gigascale AI Factories

Designed to cut east‑west communication delays in large GPU clusters, it promises lower power use with higher GPU utilization.

Overview

  • NVIDIA announced Spectrum-6, a 102.4 terabit-per-second Ethernet switch, and said deployments began on Tuesday with early installs at providers including CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla.
  • Spectrum-6 supports 200G and 800G ports, offers both pluggable and co‑packaged optics and has liquid-cooling options to reduce network power draw in very large data centers.
  • The switch is part of NVIDIA’s vertically integrated Vera Rubin and Spectrum-X stack that pairs Spectrum-6 silicon with ConnectX-9 SuperNICs and BlueField-4 DPUs to manage collective communications across hundreds of thousands of GPUs.
  • NVIDIA and partner reporting cite up to 1.6x higher AI networking performance and as much as 95% sustained network efficiency in >100,000 GPU deployments, but independent benchmarks and long-term field data are not yet public.
  • Cisco will integrate Spectrum-6 silicon into its N9100 line and earlier Spectrum-X gear saw use at Meta and Oracle in 2025, signaling an emerging ecosystem that could speed broader adoption and affect data-center power economics as AI farms scale toward gigawatt footprints.