Artificial intelligence isn’t just about algorithms. The data on which they’re trained is more important than the models themselves, some would argue, which is one of the reasons IDC predicts that more than 44 zettabytes of digital data will be created by 2020. Thankfully, the rise of big data has coincided with a continued decline in cloud storage pricing, motivated in part by cheaper media costs, better management tools, and innovations in object storage.
But not all cloud storage providers are created equal. Some lack the fine-grain management tools required to collate, process, and transfer AI model data quickly and efficiently. And not all enterprises have storage stacks optimized for data science workflows.
Nvidia and data storage company NetApp today jointly announced what they believe is a solution: Ontap AI, which they describe as an “AI-proven architecture.” Powered by Nvidia’s DGX supercomputers and NetApp’s AFF A800 cloud-connected flash storage, it’s designed to help organizations achieve “edge to core to cloud” control over their data by delivering unprecedented access and performance, said Octavian Tanase, senior vice president at NetApp.
“Our unique vision of a data pipeline [affords] simplicity of deployment,” he told VentureBeat in a phone interview. “People are looking for scale — they want to start small and grow. At the end of the day, we want customers to be able to manage data across the edge … correlate datasets, build large data lakes, [and] ultimately make faster decisions and better decisions [about] data.”
The connective tissue that ties this solution together is NetApp’s Data Fabric, a SaaS and on-premise solution that unifies data sources across clouds — specifically, clouds in datacenters, public cloud offerings from service providers, and hybrid private/public clouds — and provides lightning-fast access to data, regardless of its format or physical location.
At the core of Ontap AI is Nvidia’s DGX-1, an AI supercomputer that’s optimized for deep learning. DGX-1 sports up to 256B of GPU RAM and eight Tesla V100s Tensor Core GPUs configured in a hybrid cube-mesh topology using NVIDIA NVLink. A single DGX-1 delivers 1 petaflop of compute power — the equivalent of over 800 CPUs, Nvidia claims.
NetApp’s aforementioned AFF A800s, meanwhile, boast equally impressive performance: sub-200 microsecond latency and throughput of up to 300GB/s in a 24-node cluster.
“In the world of AI, data integration is essential,” Jim McHugh, vice president and general manager at Nvidia, said in an interview. “What’s really required for GPU AI training is quite different than traditional applications. The goal is to make it as painless as possible for data scientists, and as painless as possible for the people who build out infrastructure, too.”
One of Ontap AI’s first practitioners is Cambridge Consultants, a U.K.-based engineering consulting firm. It applied Ontap AI in a health care vertical, where it’s leveraging the tech to build systems that evaluate drug treatments and their impacts on patient outcomes. It’s also used Ontap AI to create Vincent, a deep learning program designed to learn how to paint like a human.
Other Ontap AI launch partners include IAS, Groupware Technology, ePlus, and WWT.
“Developing disruptive AI technology and turning this into breakthrough products and services for our customers is a vital requirement across many markets we work in,” Monty Barlow, head of artificial intelligence at Cambridge Consultants, said. “[It’s] simplifying and accelerating the data pipeline for deep learning.”
Nvidia’s powerful hardware platform is a big get for NetApp, which reported net revenue of $5.9 billion in May for its fiscal 2018. It’s also the only storage vendor partnered with the three largest public cloud providers — Amazon Web Services, Microsoft Azure, and Google Cloud.