There's a good reason storage providers are lining up AI systems with the likes of Nvidia. Many AI production deployments so far have been on premise.
Brian Schwarz, vice president of product management at Pure Storage, said that so far the company's production AI deployments have been on premises.
The vast majority of production deployments we've had so far have been very on premises focus. The customers have done experimentation in the cloud and then got to that point where they know it's going to run years at a time and it's strategic. The main reason people do on premises is an own vs. lease problem. A lot of the data is on premises already or generated on premises. Moving data to the cloud is a problem.
My first advice is to build your AI system where your data set is. If it's all in the cloud use the cloud. If data from a factory or hospital build the system there.
At Nvidia's GTC conference, the company is highlighting its data center momentum with partners like Pure Storage. Nvidia recently acquired Mellanox in a deal that will up the rivalry with Intel.
Specifically, Pure Storage outlined the following:
Hyperscale AIRI Nvidia DGX-1 and DGX-2 systems with Infiniband and Ethernet fabrics as interconnect options. The systems can grow across racks, said Schwarz.
Pure Storage FlashStack for AI, a hyperconverged system built with Cisco and Nvidia for AI workloads. FlashStack with AI consists of Cisco UCS C480ML and Pure Storage FlashBlade as well as Nvidia GPUs and AI libraries.
Both systems work with Nvidia's NGC software container registry and various machine learning frameworks as well as Kubernetes. Pure Storage cited Taiwan's Linkou Chang Gung Memorial Hospital, the main branch of the Chang Gung Memorial Healthcare System, and German non-profit UnternehmerTUM, an AI innovation incubator, as customers.