Nvidia releases blueprints and reference architectures for scalable AI data center design, branded as AI factories.
Nvidia Corp. announced today a series of Enterprise Reference Architectures—detailed blueprints designed to help organizations plan and build advanced, future-proof "artificial intelligence factories" to support a new generation of intelligent applications. These AI factories are high-performance and secure data centers designed to "manufacture intelligence." The challenge organizations face is that no one has built AI factories before, and with the rapid pace of development in large language model capabilities and software frameworks, best practices and standardized approaches for building such data centers are still in their infancy, making it difficult for enterprises to invest in data center infrastructure with confidence.
Nvidia's Enterprise Reference Architectures provide detailed recommendations on the full-stack hardware and software needed for AI factories, along with guidance on aspects such as the most optimal server, cluster and network configurations. According to Nvidia, by using its Enterprise RAs, companies will be able to build and deploy "cost-effective accelerated infrastructure" that's interoperable with various third-party hardware and software components, enabling it to be easily updated in the future.
The reference architectures provide recommendations for Nvidia-certified servers featuring its graphics processing units, which are the workhorse for most AI applications, as well as guidelines for AI-optimized networking using the Nvidia Spectrum-X AI Ethernet platform and the company's BlueField-3 data processing units. The blueprints also incorporate Nvidia's AI Enterprise platform, which includes microservices such as Nvidia NeMo and Nvidia NIM for building and deploying AI applications, along with Nvidia Base Command Manager Essentials, which provides tools for infrastructure provisioning, workload management and resource monitoring. The reference architectures will be made available to companies through server manufacturing partners, including Dell Technologies Inc., Hewlett-Packard Enterprise Co., Super Micro Computer Inc. and Lenovo Group Ltd., giving enterprises flexibility in choosing their underlying server platforms.
The primary benefits of using Nvidia's reference architectures include faster deployment and maximum performance. As Bob Petter, vice president and general manager of enterprise platforms at Nvidia, stated: "Enterprise RAs reduce the time and cost of deploying AI infrastructure solutions by providing a streamlined approach for building flexible and cost-effective accelerated infrastructure." The reference architectures are also designed to be future-proofed and easily upgraded as new innovations in hardware and software become available. Analyst Holger Mueller of Constellation Research Inc. noted that most enterprises simply don't have the necessary skills and experience to build AI infrastructure themselves, and that Nvidia's approach addresses this gap. Mueller said: "Nvidia plays a key role in making almost every generative AI project work, and its blueprints will make it much easier for organizations to build and upgrade their on-premises AI architectures. So long as enterprises are happy using Nvidia's chips, and many are, this is a win-win scenario. The enterprise gets to go live with their AI projects sooner, while Nvidia bags another long term customer."