The industry’s first Ethernet network platform specifically designed for AI

Easily deploy and manage powerful AI applications with the NVIDIA Spectrum™-X network platform. With the tight coupling of NVIDIA Spectrum-4 Ethernet switches and NVIDIA® Blue Field®-3 Super NICs, Spectrum-X delivers superior performance for AI, machine learning, natural language processing, and various industry applications.

To build an entirely new AI cloud, a dedicated AI network architecture is needed. AI workloads are very unique, generating a massive amount of network data flow, which significantly impacts the network.

Running these workloads on existing cloud infrastructure built on traditional Ethernet can lead to severe congestion, increased latency, and unfair bandwidth allocation. This results in performance degradation and inability to effectively utilize the system’s GPUs.

The NVIDIA Spectrum-X network platform is the world’s first Ethernet network architecture specifically built for AI workloads, offering up to twice the AI cluster performance compared to traditional Ethernet.

This end-to-end NVIDIA solution begins with the Spectrum-4 switch, which features the impressive Spectrum-4 ASIC. With NVIDIA’s innovative backing, this ASIC offers numerous groundbreaking features, setting it apart from the rest.

The Spectrum-4 switch works in close concert with the BlueField-3 DPU to ensure optimal resource utilization and efficient data transfer within AI clusters. It features enhanced RoCE capabilities such as adaptive routing, performance isolation, and congestion control, all integrated into an optimized full-stack solution.

The Spectrum-X network platform provides acceleration technology through standard Ethernet protocols, achieving the most efficient bandwidth while reducing litter and tail latency, thus maximizing AI performance.

In this real-world example, with the same performance, energy efficiency, training time, and total cost of ownership for each GPU, Spectrum-X achieved a 1.7 times performance increase.

Furthermore, using the NCCL AllReduce AI model can increase effective bandwidth by 2.5 times, resulting in faster and more predictable training.

Continuous optimization across the software stack, libraries, and operating systems ensures the best performance and interoperability of AI infrastructure.