Accelerated Ethernet Platform for Hyperscale generative AI launched by NVIDIA

NVIDIA Spectrum-4, Acceleration Software and BlueField- 3 DPUs integrated by new NVIDIA Spectrum- X networking platform. The platform has been adopted by Global Cloud Service Providers to scale out Generative AI services. 

NVIDIA has introduced NVIDIA Spectrum-X, an accelerated networking platform designed to boost the efficacy and efficiency of Ethernet-based AI clouds. NVIDIA Spectrum-X is founded on networking advancements enabled by the close integration of the NVIDIA Spectrum-4 Ethernet switch and the NVIDIA BlueField-3 DPU. This enables 1.7x greater overall AI performance and power efficiency, along with consistent, predictable performance in multi-tenant environments.

Spectrum-X is augmented with NVIDIA acceleration software and software development kits (SDKs), enabling developers to create software-defined, cloud-native AI applications. When end-to-end capabilities are being delivered, massive transformer-based generative AI models are able to function at a higher speed. This enables cloud service providers, network engineers, and AI data scientists to improve their results and make decisions more swiftly.

NVIDIA Spectrum-X is being used by some of the most prominent hyperscalers in the world, including some of the most innovative cloud pioneers in the business. NVIDIA is building Israel-1that is  a hyperscale generative AI supercomputer that will be installed in its Israeli data center on NVIDIA HGXTM H100 eight-GPU-based Dell PowerEdge XE9680 servers, BlueField-3 DPUs, and Spectrum-4 switches. Israel-1 will serve as a blueprint and testbed for NVIDIA Spectrum-X reference designs.

According to Gilad Shainer, senior vice president of networking at NVIDIA, transformative technologies like generative AI are compelling businesses to stretch the limits of data center performance in pursuit of a competitive advantage. NVIDIA Spectrum-X is a new category of Ethernet networking that overcomes barriers for AI workloads of the next generation that have the potential to disrupt entire industries. These AI tasks have the capacity to revolutionize entire industries.

As a result of its high degree of adaptability, the NVIDIA Spectrum-X networking platform is suitable for use in a wide variety of AI-related endeavors. It utilizes Ethernet that conforms to industry standards and is compatible with other systems that use Ethernet.

Spectrum-4, the world’s first 51Tb/sec Ethernet switch designed specifically for AI networks, is the foundation of the platform. Advanced RoCE extensions combined across Spectrum-4 switches, BlueField-3 DPUs, and NVIDIA LinkX optics to create a 400GbE AI-optimized network from end to end.

NVIDIA Spectrum-X improves multitenancy. This is done by isolating performance to ensure that AI workloads operate optimally and consistently for each tenant. In addition to this, it offers improved visibility into AI performance by locating performance bottlenecks and automating fabric validation.

Acceleration software powering Spectrum-X consists of potent NVIDIA SDKs, such as Cumulus Linux, pure SONiC, and NetQ, which facilitate the networking platform’s extreme performance. In addition, it contains the NVIDIA DOCA software framework, which is the core of BlueField DPUs.

In order to support the growth and extension of AI clouds while maintaining high performance levels and minimizing network latency, NVIDIA Spectrum-X provides an unparalleled scale of 256 200Gb/s ports connected by a single switch or 16,000 terminals in a leaf-spine topology with two tiers. These port counts can be reached by connecting a single switch.