NVIDIA announced the Enterprise Reference Architectures (Enterprise RA) initiative, within which partners and customers will be able to use reference architectures to build their own enterprise-grade AI platforms designed for resource-intensive workloads.

NVIDIA notes that as enterprises move from general-purpose computing to accelerated computing, they face various challenges when designing and deploying data center infrastructure. This makes it difficult to develop long-term strategies and reduces the effectiveness of investments. The Enterprise RA initiative aims to solve the problems.

Image source: NVIDIA

Enterprise RA reference architectures will help organizations minimize design errors in so-called AI factories (data centers for AI workloads) by providing comprehensive hardware and software recommendations, as well as detailed guidance on optimal server, cluster, and network configurations. As a result, customers will be able to reduce costs and reduce the time required to build the next generation of AI computing infrastructure.

The advantages of the proposed approach include scalability and manageability, a high level of security (zero trust principle is applied), optimal performance, reduced system complexity and accelerated time to market. The reference architectures are designed to be easily upgradeable in the future. Compatibility with various third-party hardware and software components is mentioned, but the list of recommendations primarily includes solutions from NVIDIA itself, including:

  • Certified servers with AI accelerators based on NVIDIA GPUs;
  • Optimized network platform based on NVIDIA Spectrum-X AI Ethernet and NVIDIA BlueField-3 DPU;
  • NVIDIA AI Enterprise software components, including NVIDIA NeMo and NVIDIA NIM microservices for quickly building and deploying AI applications.

Solutions based on NVIDIA Enterprise RA will be offered by NVIDIA partners, including Dell, HPE, Lenovo and Supermicro.

Leave a Reply

Your email address will not be published. Required fields are marked *