Nvidia has introduced its Spectrum-6 Ethernet switch, a networking chip designed to handle the massive data loads of gigascale AI factories. The switch delivers 102.4 Tb/s of switching capacity, making it one of the fastest Ethernet switches on the market.
Built for AI Workloads
The Spectrum-6 is purpose-built for the kind of distributed computing that powers large language models and other AI training clusters. Unlike traditional data center switches, it prioritizes low latency and high throughput for GPU-to-GPU communication. Nvidia says the switch can support up to 512 ports of 400 Gb/s Ethernet, or a mix of slower and faster links, giving operators flexibility in designing their networks.
Who's Already on Board
Meta, Oracle, Cisco, and Nebius are among the early partners or customers for the Spectrum-6. Meta plans to use the switch in its AI research infrastructure, while Oracle is integrating it into its cloud services. Cisco, a longtime networking rival, is also working with Nvidia on interoperability. Nebius, a European cloud provider, will deploy the switch in its GPU clusters.
Why Ethernet for AI?
Ethernet has long been the standard for general-purpose networking, but AI workloads have increasingly pushed toward specialized interconnects like Nvidia's own InfiniBand. With Spectrum-6, Nvidia is betting that Ethernet can match InfiniBand's performance for AI while offering easier integration with existing data center equipment. The company claims the switch delivers up to 3.5x the performance of previous-generation Ethernet switches for AI workloads.
The switch is expected to begin shipping to customers in the second half of 2025. Pricing has not been disclosed.
