NVIDIA releases new switches to comprehensively optimize trillion parameter level GPU computing and AI infrastructure

On March 19, Beijing time, NVIDIA released a new network switch, the X800 series, specially designed for large-scale AI during GTC 2024.

NVIDIA Quantum X800 InfiniBand network and NVIDIA Spectrum X800 Ethernet network are the world's first network platforms with end-to-end throughput of up to 800Gb/s, which have raised the network performance of computing and AI workloads to a new level and matched them Software The combination of strong players can further accelerate AI, cloud, data processing and high-performance computing (HPC) applications in various data centers, including data centers based on the latest NVIDIA Blackwell architecture products.

Gilad Shainer, Senior Vice President of NVIDIA Network, said: "NVIDIA network platform is the core of AI supercomputing infrastructure to achieve large-scale scalability. The end-to-end network platform based on NVIDIA X800 switch makes the new AI infrastructure easily run trillion parameter level generated AI services."

Early users of Quantum InfiniBand and Spectra-X Ethernet include Microsoft Azure, Oracle Cloud Infrastructure and Coreweave.

Nidhi Chappell, vice president of infrastructure of Microsoft Azure AI, said: "AI is a powerful tool to transform data into knowledge. The foundation supporting this transformation is that the data center has evolved into a high-performance AI engine, which has put forward greater demands on network infrastructure. Through the new integration with NVIDIA network solutions, Microsoft Azure Cloud will continue to build our infrastructure and promote AI Cloud to a new level. "

Set a new benchmark for ultimate performance

The Quantum X800 platform has set a new benchmark for the ultimate performance of AI dedicated infrastructure. The platform includes the NVIDIA Quantum Q3400 switch and NVIDIA ConnectX-8 SuperNIC. The interconnection of the two has reached the industry-leading end-to-end 800Gb/s throughput. The switching bandwidth capacity has been increased by 5 times compared with the previous generation of products. The network computing capacity has been increased by 9 times with NVIDIA SHARP technology (SHARPv4), It reached 14.4Tflops.

Spectrum X800 platform brings optimized network performance to AI cloud and enterprise infrastructure. With the 800Gb/s Spectrum SN5600 switch and NVIDIA BlueField-3 SuperNIC, the Spectrum X800 platform provides various critical advanced functions for multi tenant generative AI clouds and large enterprise users.

Spectrum X800 accelerates the processing, analysis and execution of AI workloads by optimizing network performance, thus shortening the development, deployment and time to market of AI solutions. The Spectrum X800 is specially designed for multi tenant environments, which realizes the performance isolation of AI workloads of each tenant, keeps business performance in the best state, and improves customer satisfaction and service quality.

NVIDIA software helps

NVIDIA provides network acceleration communication library, software development kit and Management software And other software solutions.

The NVIDIA integrated communication library (NCCL) can extend the parallel computing tasks of GPU to the Quantum-X800 network, and use its strong network computing capability based on SHARPv4 and support for FP8 to provide super performance for large model training and generative AI.

NVIDIA's full stack software solution brings advanced programmability, making the data center network more flexible, reliable and sensitive, which not only improves the overall operational efficiency, but also meets the needs of modern applications and services.

Ecological milestone

Several leading infrastructure suppliers and system manufacturers around the world will start to provide network platforms based on Quantum X800 and Spectrum X800 next year, including Aivres, DDN, Dell Technologies Eviden、Hitachi Vantara、 Huihe, Lenovo, AMD, VAST Data, etc.

This article is an original article. If it is reproduced, please indicate the source: NVIDIA releases new switches to comprehensively optimize trillion parameter level GPU computing and AI infrastructure https://net.zol.com.cn/861/8613009.html