As the world becomes increasingly data-driven, AI/ML algorithms are being incorporated in most business applications. Historically, data in AI architectures was moved to a central location to perform both model training and inference. This centralized approach is becoming untenable due to cost, performance, and privacy reasons.
During his talk “The Era of Distributed AI Architectures,” Kaladhar Voruganti, senior fellow, technology and architecture, office of the CTO, Equinix, shared his thoughts on next-generation distributed AI architectures.
“AI has been around for a while,” Voruganti said. “Once we started collecting a lot of data, previous techniques were slow and not really accurate.”
Currently we are in the present era where companies aren’t realizing that they don’t have the data that they need and wind up pulling tons of information from outside their organizations. The size of datasets at the edge is becoming huge, he explained.
To make insights available to subject matter experts, companies want operational simplicity. AI model building and model inferencing can help. Examples of this include image recognition, voice recognition, churn prediction, and fraud detection.
AI is becoming distributed, there are many types of edges in play, he noted. There is the far edge, micro edge, and macro edge where information moves and can connect to other data points.
“A lot of 5G deployments will be at the far edge,” he said. “The same application may have microservices and capabilities at any of those edges at the same time.”
There are three types of distributed AI models: compute to data, compute and data to a neutural location, data to capture.
There are several governance models as well. This includes the consortium member driven model, founder driven model, and a marketplace operator driven model. The location of AI stack matters for all of the distributed architectures, he said.
The annual Data Summit conference returned in-person to Boston, May 17-18, 2022, with pre-conference workshops on May 16.
Many Data Summit 2022 presentations are available for review at https://www.dbta.com/DataSummit/2022/Presentations.aspx.