Artificial intelligence models are growing more complex and data-hungry. Across the industry, organizations are rethinking how they manage and refine data to keep pace with accelerated computing and real-time intelligence needs, driving a new focus on AI-ready storage infrastructure.

These days, many enterprises are trying to pour as much power as possible into massive graphics processing unit clusters. The goal is to run AI training efficiently at massive scale and, ultimately, influence what’s possible with AI in the future, according to Alan Bumgarner (pictured, right), director of strategic planning for the Data Center Group and AI technologist at Solidigm, a trademark of SK Hynix NAND Products Solutions Corp.

“When you break down the mechanics of how all of this stuff really operates inside of a data center, it really turns your data center into a warehouse-style computer,” Bumgarner said. “When you have a very large graphics cluster … you can’t feed that air, so you have to feed it data.”

Bumgarner and Glenn Dekhayser (left), global principal technologist at Equinix Inc., spoke with theCUBE’s Dave Vellante and Savannah Peterson at SC25, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They explored how well-organized data underpins effective AI and how the industry is moving toward AI-ready storage infrastructure. (* Disclosure below.)

AI-ready storage infrastructure in focus

GPUs need data pulled in from a large network tied to clean, organized storage. Once that data is ingested, the system can drive through model layers in an extremely latency-sensitive process where more power and more data mean faster training and better accuracy, according to Bumgarner.

“The less power you have to [have] … and the more efficient it becomes, naturally, the better the calculations become, the more accurate your models become and all of those things happen,” Bumgarner said. “It’s a very close relationship between what your storage array can do to keep your data clean and in a proper global namespace to pull it over to make this model more accurate — to do the things that you were trying to achieve.”

Many enterprises initially tried to build their own on-premises AI stacks to avoid costly data movement and maintain control, but soon discovered that retrofitting traditional data centers for modern AI wasn’t practical. The emerging reality is that AI success requires an “and, not or” strategy, according to Dekhayser.

“You’re not going to go in the cloud or on-prem or to a neocloud or to the edge. Every use case you [have] will look different,” Dekhayser said. “You might use a different model, which may imply a different provider, which may require a different network, different data. So what we’re seeing enterprises start to coalesce around is some best practices.”

Because each AI use case demands something slightly different, organizations will need to combine multiple environments rather than rely on just one, according to Dekhayser.  Many will start proofs of concept with hyperscaler GPUs or shift pilots to neocloud platforms — and potentially run production there as needs evolve.

“But the one thing that’s repetitive across all these use cases is that you need to acquire the data. Where are you getting that from?” Dekhayser said. “All of your edges, all of your [internet of things] devices, all your logs, all your customers. You got to acquire all that data, you got to manage that data in its raw, ugly format.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of SC25:

(* Disclosure: Solidigm sponsored this segment of theCUBE. Neither Solidigm nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE