NVIDIA introduced Spectrum-6, a 102.4-terabit-per-second Ethernet switch system delivering double the capacity of previo
NVIDIA Spectrum-6 marks a networking milestone in AI infrastructure, delivering a 102.4-terabit-per-second Ethernet switch system that provides twice the capacity of prior-generation systems. The technology is built as part of the NVIDIA Vera Rubin platform and anchors the next generation of the NVIDIA Spectrum-X Ethernet platform, engineered to deliver the bandwidth, scale and intelligence required to operate an AI factory as a unified end-to-end computing system.
The world's leading AI infrastructure builders, including CoreWeave, Microsoft, Nebius, SpaceXAI and Tesla, will be among the first to deploy Spectrum-6 to accelerate their AI factories. CoreWeave's director of product for networking, Min Jun, stated that networking is central to delivering performance at scale and that bringing Spectrum-6 and liquid-cooled Spectrum-X Ethernet infrastructure into their AI factories will help deliver the bandwidth, resilience and efficiency needed to train frontier models and deploy inference faster. Nebius vice president Laurelle Roseman emphasized that at gigascale, performance depends on coordination and keeping every GPU synchronized, which Spectrum-6 achieves through a fabric that remains fast and resilient as demanding workloads scale.
The platform achieves its performance through a fundamental architectural shift. Large-scale AI training and inference depend on thousands of accelerators exchanging data continuously through collective communications that generate intense east-west traffic. Traditional Ethernet was designed for enterprise applications and north-south traffic patterns between users, servers and storage, not the synchronized collective-heavy communication patterns required at gigascale AI. Spectrum-X Ethernet transforms Ethernet into a purpose-built, high-performance scale-out fabric engineered to keep every GPU fed with data.
Spectrum-6 combines with the NVIDIA ConnectX-9 SuperNIC to form the next generation of Spectrum-X Ethernet and is engineered as part of NVIDIA Vera Rubin, integrating the NVIDIA Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch. The technology supports both pluggable and co-packaged optics, and offerings include liquid cooling to enable comprehensive end-to-end cooling for the entire AI factory while increasing network power efficiency.
The NVIDIA Spectrum-X Ethernet networking platform for AI combines intelligent switches, ConnectX-9 SuperNICs and full-stack networking software engineered together for AI workloads. Advanced features continuously optimize traffic movement through the fabric, intelligently balancing traffic across available paths, rapidly bypassing failures and precisely recovering when data fails to reach its destination. The platform supports open network operating systems and offers a choice of RDMA transport models, providing AI builders flexibility without compromising performance.
Spectrum-X Ethernet delivers up to 1.6 times higher AI networking performance than off-the-shelf Ethernet and sustains up to 95 percent network efficiency across deployments exceeding 100,000 GPUs. Hardware-accelerated Spectrum-X multiplane topologies reduce the number of switches required in data centers by 1.7 times. When layered with Spectrum-X Ethernet Photonics, network fabric enhancements achieve 5 times higher power efficiency and 10 times improved mean time between incidents, delivering workload accelerations beyond component-level optimization.