NVIDIA Enterprise Reference Architecture (Enterprise RA)

What is NVIDIA Enterprise RA

TheNVIDIA Enterprise Reference Architecture (Enterprise RA) is a comprehensive set of proven guidelines and best practices designed to provide a scalable, productive, and secure infrastructure for enterprise AI solutions. It covers hardware, software, and optimal server, cluster, and network configurations for today's AI business challenges.

Purpose:

• Reduce complexity when designing and deploying data center infrastructure.
• Enterprise RA provides proven and comprehensive design recommendations for large-scale deployment of systems such as the H200 NVL.
• Enterprise RA helps accelerate time to market for partners and customers building data center solutions

Composition

• At the heart of each Enterprise RA is an optimized NVIDIA System certified server that follows a prescribed design pattern to ensure optimal performance when deployed in a clustered environment.
• There are different types of server configurations for which Enterprise RAs are designed, including PCIe Optimized 2-4-3, PCIe Optimized 2-8-5, and HGX systems. Numerical designations, such as in "2-8-5," indicate the number of sockets (processors), the number of graphics processing units (GPUs), and the number of network adapters, respectively.
• For example, the Enterprise RA for the H200 NVL uses the PCIe Optimized 2-8-5 reference configuration. This configuration reduces latency, reduces CPU utilization, and increases network bandwidth for real-time operations, which is critical for efficient data processing.

NVIDIA Technology Integration

• Enterprise RAs include recommendations for using the NVIDIA Spectrum-X Ethernet platform to maximize performance when deploying AI systems in a clustered environment. This includes Spectrum-4 switches and SuperNIC BlueField-3 network adapters.
• For peak network performance, Enterprise RA recommends a dedicated BlueField-3 SuperNIC with 400 Gbps connectivity for every two H200 NVL GPUs in the cluster.
• Also, Enterprise RA for the H200 NVL utilizes the NVIDIA Collective Communications Library (NCCL) to provide efficient, low-latency communication and scalability for workloads that require efficient communication between multiple GPUs.

Application:

• Suitable for enterprise data centers, clouds, real-time data transfer, autonomous driving solutions, and big data analytics, as well as for building AI factories focused on generative AI and large language models (LLMs).
• Dell Technologies implements Enterprise RA in its clusters based on PowerEdge servers (e.g., R760xa with a 2-4-3 configuration and XE9680 with 2-8-9), demonstrating the industrial application of the architecture.
• Software Stack: Powered by NVIDIA AI Enterprise, which includes drivers, GPU management tools for Kubernetes (GPU Operator), networking (Network Operator), AI microservices (NeMo, NIM), and infrastructure management (Base Command Manager).