Nvidia launched its Spectrum-6 switch, a 102.4 Tb/s system the chip giant claims offers twice the capacity of previous-generation systems.
Set to form the networking anchor for Nvidia’s upcoming Vera Rubin platform, Spectrum-6 is Ethernet-based and, per an earlier blog post, the SN6810 version features 128 ports at 800 Gb/s. There’s also the SN6800, which the self-proclaimed “largest networking company in the world” said offers a truly staggering 409.6 Tb/s of bandwidth across 512 ports at 800 Gb/s. Both are liquid-cooled.
Microsoft, CoreWeave, Nebius, and SpaceXAI are listed among the initial AI infrastructure operators set to deploy the switch lines, with Nvidia touting it as a means to provide more compute capacity and ultimately help customers train models and deploy inference services far faster.
Scot Schultz, Nvidia’s senior director for high-performance computing (HPC) and technical computing, said Spectrum-6 ensures higher utilization for AI operators during demand peaks, with Nvidia’s underlying Spectrum-X fabric providing the switch line with up to 1.6-times higher AI networking performance than off-the-shelf Ethernet lines, sustaining up to 95% network efficiency across deployments surpassing 100,000 graphic processing units (GPUs).
“At gigascale, performance comes down to coordination: keeping every GPU in lockstep so one slow link doesn’t stall an entire job,” Laurelle Roseman, VP of global partnerships at Nebius, noted. “That’s what Nvidia Spectrum-6 goes after and why we brought it in early – a fabric that stays fast and resilient as our customers’ most demanding workloads scale.”
New to Nvidia’s Spectrum switch line is support for both pluggable and co-packaged optics (CPO) form factors. The latter line has been highly anticipated for some time and is expected to start shipping later this year. Several lucky customers have already been testing the CPO switches, with neocloud darling Lambda showing off the InfiniBand-based Q3450-LD in early June. Schultz’s Nvidia blog claims the addition of Spectrum-X Ethernet photonics provides its new switch line with five-times higher power efficiency and 10-times improved mean time between incidents.
“Bringing Nvidia Spectrum-6 and liquid-cooled Spectrum-X Ethernet infrastructure into our AI factories will help us deliver the bandwidth, resilience and efficiency customers need to train frontier models and deploy inference faster,” Min Jun, director of product for networking at CoreWeave, added.