NVIDIA announced the launch of the new NVIDIA Spectrum-X platform, which is designed to increase the performance and effectiveness of cloud solutions based on Ethernet using artificial intelligence (AI). The NVIDIA Spectrum-X platform combines the Ethernet-Commutator NVIDIA Spectrum-4, DPU (Data Processing Unit) Nvidia Bluefield-3 and special software for accelerating AI applications. Thanks to this, the platform reaches almost half the best indicators in the overall performance and energy efficiency of AI, and also provides stable and predictable work in multi -user environments.
NVIDIA Spectrum-X platform supports Ethernet standards and is compatible with existing Ethernet glasses. It can be used in various AI applications, such as generative models of the language (GPT and BERT), distributed learning and parallel processing, processing of a natural language (NLP), computer vision, high-performance modeling (NVIDIA OMNIVERSE and NVIDIA OVX) and data analytics (spark ). The platform begins with the spectrum-4 switch, which is the first Ethernet-community, created specifically for AI networks. It has a capacity of 51 Tbit/s and can connect 128 ports of 400 Gb/s or 64 ports 800 Gb/s. The switch is paired with the DPU Bluefield-3, which is responsible for the transmission and priority of data, as well as with LinkX optical transceivers, which create an optimized network of 400GBE.
NVIDIA Spectrum-X platform also improves multiplayer work due to performance isolation, guaranteeing the optimal and constant operation of the AI-user for different customers. It can determine narrow places in performance and has an automatic testing of the network state to ensure better performance of AI performance.
To accelerate the operation of the NVIDIA Spectrum-X platform, it uses the software and tools for the developer (SDK) from NVIDIA, such as Cumulus Linux, Pure Sonic, Netq and Doca. They allow the platform to achieve extreme performance.
As a sample and test stand for the reference designs NVIDIA Spectrum-X, NVIDIA is building a hyper-captive ISRAEL-1 hyper-controlle, which will be deployed in its Israeli Data Center on the Dell Poweredge XE9680 servers on the basis of eight-GPU NVIDIA HGX H100 with DPU Bluefield-3 and spectrum-4 switches.
Among companies that offer solutions based on Nvidia Spectrum-X, you can call Dell Technologies, Lenovo and Supermicro.
NVIDIA Spectrum-X is part of the NVIDIA strategy to create program-defined AI-date-centers of a new generation. Previously, the company has already presented its first processor for GRACE data processing centers, which is also optimized for working with AI-USA. In addition, the company continues to develop its GPU line for AI, including the new Grace Hopper chip, which has already entered full production.
NVIDIA also actively cooperates with other leaders of the development industry. For example, it united with Google Cloud to create Vertex AI, a single platform for the development and management of AI models. And together with Microsoft, she launched the Azure ARC service, which allows you to deploy AI applications on any devices with DPU Bluefield.