Artificial intelligence is a key transformation driver across industries, and there’s a race to unlock its full potential. For that to happen, key players in the tech world are racing to build the infrastructure needed to sustain its momentum. These efforts span AI accelerators, system architectures, data centers, networking and storage.
“AI is driving a huge appetite for performance, and everybody is constrained by budgets and space and power,” said Steve Scott (pictured), corporate fellow, network and systems architecture at Advanced Micro Devices Inc. “The COLA rates are historically low, like less than 2%. I’ve heard that over 80% of new data centers that are under construction are already leased. One of the things that customers are asking us about is just the need to consolidate their general-purpose compute to free up space and power for AI.”
Scott spoke with Raj Yavatkar, chief technology officer of Juniper Networks Inc., during a panel discussion at the Seize the AI Moment event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. He was joined by Justin Hotard, executive vice president and general manager of data center and AI Business Group at Intel Corp.; Andrew Feldman, founder and chief executive officer of Cerebras Systems Inc.; and Shimon Ben-David, chief technology officer of WekaIO Inc. They discussed the importance of specialized AI accelerators, flexible infrastructure and open ecosystems in today’s enterprise. (* Disclosure below.)
AI accelerators and open ecosystems
Traditional GPU architecture, designed to render pixels on a screen, wasn’t inherently optimized for AI workloads. This observation led to the development of new, AI-optimized machines capable of reducing training times—one of the primary bottlenecks in AI development, according to Feldman.
“We thought that there would be an opportunity to build a dedicated machine optimized for AI,” he said. “We thought for sure that we could build a better machine than a wide vector SIMD processor with limited memory bandwidth, which is the family that the graphics processing unit lives in. We spoke to dozens of customers and everybody was having trouble with training times.”
Optimizing for AI workloads often involves consolidating older servers to free up space and power. Intel’s Epic CPUs are helping customers achieve significant consolidation, reducing the number of physical servers required and lowering energy consumption, Hotard explained.
“One of the things we’re doing at Intel is trying to provide lower cost access points, things like our developer cloud where we can give access to customers to get started to learn to test some of the ideas and tools they have,” he said. “The second thing I would say is it’s all about leveraging the data that enterprises have. Over 80% of the world’s data is in enterprises. The majority of that proprietary compute has generated so much data over the last few decades.”
As AI continues to evolve, so do the storage requirements that support AI and machine learning workloads. AI models require vast amounts of data, often accumulating from diverse sources such as fleets of data brokers or HPC simulations. Traditional storage systems, designed for enterprise or HPC workloads, can’t keep pace with the unique demands of AI storage environments, according to Ben-David.
“The interesting thing is that every stage has its own IO pattern,” he said. “There’s massive requirements for storage or data environments in AI that behave completely differently than traditional enterprise storage or traditional HPC storage. It’s a new thing completely.”
Here’s the complete video interview, part of SiliconANGLE’s and theCUBE Research’s coverage of the Seize the AI Moment event:
Here’s the complete event video playlist:
(* Disclosure: TheCUBE is a paid media partner for the Seize the AI Moment event. Neither Juniper Networks Inc., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)