Nvidia, VMware Team Up to Virtualize AI Compute

Nvidia borrowed the “build it and they will come” philosophy to driving the adoption of artificial intelligence (AI) today with the launch of the AI Enterprise software suite. Nvidia borrowed the “build it and they will come” philosophy to driving the adoption of artificial intelligence (AI) today with the launch of the AI Enterprise software suite. The suite is packed to the gills with tools and frameworks for building AI and machine learning applications. It launches alongside an update to VMware’s vSphere 7, which extends support for AI workloads built using Nvidia’s suite into a virtualized environment. According to Nvidia, AI Enterprise provides customers with everything they need to develop AI applications for healthcare, manufacturing, and financial services to name a few, while the integration with vSphere means they can deploy those workloads using the same tools they’re already using to manage their data center infrastructure. “This combination enables scale-out, multi-node performance, and compatibility for a vast set of accelerated CUDA applications, AI frameworks, models, and SDKs for the hundreds of thousands of enterprises that use vSphere and server virtualization,” said Justin Boitano, VP and GM of enterprise and edge computing at Nvidia. AI workloads have traditionally run on bare-metal servers, he explained, adding that customers can now deploy AI workloads in a virtual environment, thus reducing time to deployment from 80 weeks to just eight. Under vSphere, AI workloads can now be spread across multiple nodes without a substantive performance hit compared to bare metal, Nvidia claimed. Meanwhile, enterprises with smaller AI workloads can now run multiple models on a single Nvidia A100 GPU thanks to support for Multi-Instance GPU (MIG) in vSphere. MIG allows a single GPU to be subdivided into seven logical GPUs that can be separately addressed. To date, vSphere is the only hypervisor with support for live migration of MIG deployments. In addition to being able to run on existing infrastructure, Nvidia’s software suite is optimized for use on Nvidia A100 tensor-core GPUs running in systems from Dell Technologies, Hewlett Packard Enterprise, Lenovo, and Supermicro. Nvidia has a long history of building ecosystems around its hardware and software products. “If there is anything Nvidia has been really good at over the years, it is putting together systems in order to speed [up and] reduce the complexity of deploying their technology,” said Zeus Kerravala, principal analyst at ZK Research. “If you think about what it takes to deploy an AI system, there’s a lot of stuff.” With the introduction of vSphere 7 support, Kerravala said VMware and Nvidia have once again lowered the barrier to running AI models. “The whole idea here is to give the customers all the things they need in a pre-tuned and pre-configured way, so they can start to spend more time doing AI and less time tweaking and tuning hardware and software,” he said. Kerravala credits this end-to-end philosophy for much of Nvidia’s success in the GPU market.