While the latest industrial-grade hardware from NVIDIA isn’t about GPU, it is still about how to make AI computing better in effectiveness and efficiency, hence, the new Team Green announcement is made for the Spectrum-XGS Ethernet.
This new networking breakthrough is created to connect distributed data centers to form “giga-scale AI super-factories” that are usually restricted by the limits of power and capacity, as Spectrum-XGS fixes this by introducing a “scale-across” infrastructure – essentially a third pillar of AI computing that goes beyond the typical scale-up and scale-out approaches. It allows multiple facilities across cities, nations, or even continents to function as one massive AI cluster.
Built into NVIDIA’s Spectrum-X platform and comes with adaptive algorithms that handle distance between sites, it also features advanced congestion control and latency management, resulting in performance that is nearly doubled from the Collective Communications Library, speeding up multi-GPU and multi-node communication across huge distances.
CoreWeave is one of the first to embrace the technology, with cofounder and CTO Peter Salanki saying Spectrum-XGS lets them unify data centers into a single, powerful supercomputer that gives customers “access to giga-scale AI” for accelerating breakthroughs across industries.
Spectrum-XGS Ethernet joins NVIDIA’s growing portfolio of networking innovations, sitting alongside Spectrum-X and Quantum-X photonics-based switches that aim to connect millions of GPUs while reducing energy use and costs.
The system is already available as part of the Spectrum-X Ethernet platform, opening the door for hyperscale enterprises and AI pioneers to scale far beyond a single building.
