Nvidia shows a fast and reliable alternative to classic Ethernet switches. Spectrum-X works in tandem with Nvidia’s DPUs and aims to bring performance reminiscent of an HPC connection to the Ethernet market.

Nvidia wants faster and more reliable connections over Ethernet. To this end, the company is launching Spectrum-X. This is a family of switches tailored for AI environments that never drop packets. Spectrum-4 is the first device in the family. The thing delivers a throughput of 51 Tb/s and can route 128 ports for 400 Gigabit Ethernet (or 64 ports at twice the speed). According to Nvidia, the Switch is ideal for what they call AI networking.

For large data centers

The Ethernet switch is based on a huge microchip, which is manufactured at TSMC in 4 nm and enables a TDP of 500 watts. CEO Jensen Huang expects the hardware to better connect AI data centers. Spectrum-X can better manage and control traffic on a network to avoid congestion. The hardware is primarily intended for large data centers, for example in hyperscalers, where very powerful hardware is available. Think of servers running Hopper H100 accelerators, which are part of the AI ​​training infrastructure.

With Spectrum-X, Nvidia is targeting Broadcom, which is the de facto main supplier of microchips in switches. Nvidia hopes to take the top of the segment with extreme performance. With the Nvidia switches it is possible to connect servers with low latency and high reliability, similar to a dedicated connection.