Nvidia has announced its new networking platform called Spectrum-X, designed specifically for generative AI workloads. The platform utilizes the Spectrum-4 Ethernet switch and BlueField-3 DPU, promising lightning-speed lossless networking and increased performance and power efficiency.

According to Nvidia, Spectrum-X reduces run-times of massive transformer-based generative AI models, enabling companies to make faster decisions and unlock cost-saving potential. The Spectrum-4 Ethernet switch supports speeds of 51 Tb/sec, enabling a 400 Gigabit Ethernet network with Nvidia LinkX optics.

However, Broadcom’s SVP of the core switching group, Ram Velaga, claims that Nvidia’s Spectrum-X is not a unique concept and that Broadcom’s Tomahawk5 and Jericho3-AI switch ASICs already provide similar capabilities. Velaga sees Nvidia’s move as an admission that Ethernet is more suitable for handling GPU flows in AI.

Similar Post

Nvidia’s Spectrum-X platform, although not a product itself, consists of hardware and software, including the Spectrum-4 Ethernet switch and BlueField-3 DPU. It aims to mitigate traffic congestion and eliminate packet loss, leading to a highly efficient Ethernet fabric.

While Broadcom argues that the idea of “lossless Ethernet” is just marketing, Nvidia claims that Spectrum-X offers capabilities specifically designed for AI workloads. Nvidia’s CEO, Jensen Huang, showcased the Spectrum-4 switch chip and highlighted its design during a presentation at the Computex expo.

Broadcom believes that Nvidia’s vertical integration approach conflicts with the open ecosystem of Ethernet, potentially making it a tough sell for cloud providers who prefer vendor-agnostic solutions. Nvidia’s Spectrum-X relies on having both the Spectrum-4 switch and BlueField-3 DPU, making it difficult to swap components without losing features.

Nvidia’s Gilad Shainer suggests that cloud providers can still use their existing DPUs for infrastructure management and north/south traffic while using Nvidia’s BlueField-3 for east/west traffic within the cluster. He also states that Nvidia’s switches and DPUs can be deployed as standalone products.

It remains to be seen how customers will respond to Nvidia’s vertically integrated Ethernet solution in a disaggregated world where openness and vendor neutrality are valued.