AI represents a new approach to turning unstructured data into insights. It holds tremendous promise for enterprises looking to innovate faster and extend their competitive edge. As such, investment in AI initiatives has skyrocketed, with companies worldwide investing between $26B and $39B in 2016 alone, according to a McKinsey report.
However, AI requires a different approach to scale-out infrastructure and its complexities are holding back enterprises from moving forward towards an AI-first world. Indeed, deep neural networks comprise millions of trainable parameters, connected through a configurable network of layers. In addition, each network has countless variations of topologies, connection strategies, learning rates, etc. – collectively referred to as hyper-parameters.
At Big Data LDN, Brian Carpenter, senior director, technology strategy, Pure Storage, explained how enterprises could seamlessly fit AI projects into a unified analytics plan with the right infrastructure built for scaling.
“AI software is only as good as the hardware that it runs on,” said Carpenter. “While data scientists don’t have to be infrastructure experts, infrastructure determines their success.”
Why you need to modernise
Carpenter argued that because AI requires new, modern technologies like GPUs, scale-out flash, and RDMA fabric to move tremendous amounts of data, too many AI initiatives are stalled with the complexities of a ‘do-it-yourself’ approach using legacy technologies, leading to months of delays and idle time.
Indeed, from an infrastructure perspective, one thing is clear. As AI projects grow beyond the experimentation stage to adoption, demand for more computing resources is inevitably going to increase.
In addition to large-scale computational requirements, recent research shows that training accuracy increases logarithmically with the volume of training data. Thus, small improvements in accuracy could require a 10x or greater increase in dataset size. Along with large-scale computing power, creating state-of-the-art models requires larger, more diverse data sets with high-performance access.
The truth about digital transformation – Cloud and on-premise: the best of both worlds
What does AI-ready infrastructure look like?
According to Carpenter, AI-ready infrastructure should be:
- Scalable: Can’t bring everything down whenever one application needs more compute.
- Flexible: This is from a tools perspective; tools change every day. Enterprises need to avoid lock-in to keep pace with rapidly evolving ecosystems and the rapidly changing needs of data science team as their work matures.
- Portable: You want to be able to take things on and off the cloud. Think of experimental vs. productive environments.
Carpenter added: “If you have an 18-month roadmap for improving your infrastructure, you’re not moving quickly enough. If your tools change every six months, you’re crazy to go past 12 months.”
While you don’t want to set lines in the sand, according to Carpenter, you want to have a general idea of where you’re going.
The data journey: It’s only the beginning for digital transformation — Big Data LDN
