AI cluster technology drives scalable performance

Artificial intelligence is undergoing a major transformation as AI cluster technology advances, revolutionizing how industries integrate and implement this powerful tool.

This emerging technology is not only speeding up AI processes but also making them more efficient and cost-effective, driving broader adoption and accessibility across various sectors, said David Kanter (pictured, middle), founder and head of MLPerf at MLCommons Association.

“One of the things that I’m excited about that the team has done recently is first of all, we’ve added a lot of gen AI benchmarks,” Kanter said. “Then we also added power measurement so that you can see whether it’s data center inference or training of these large-scale models, how much power and energy are you using? And we’ve seen in the five years that we’ve been around, we were able to get something like 50 times better performance, which is way faster than what we would expect.”

Kanter and Adi Gangidi (right), RDMA systems for AI training at  Meta Platforms Inc., spoke with the host Rakesh Kumar (left), senior engineering leader at Juniper Networks Inc. at the Seize the AI Moment event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how AI cluster technology and collaborative benchmarking efforts are driving the next phase of AI adoption by enhancing infrastructure efficiency, scalability and performance. (* Disclosure below.)

Best practices in AI cluster technology for scalable performance

One of the key drivers behind this exponential growth in AI performance is collaboration between organizations, academia and engineers. MLCommons, a nonprofit industry consortium, exemplifies this approach by bringing together diverse players to create standardized benchmarks that measure AI performance, Kanter added.

“So MLCommons really got started with the MLPerf benchmarks as part of what’s bringing us all together today,” he said. “And it was sort of founded in the early days of machine learning and we didn’t have good standard ways of measuring performance. And so, we got the whole community together. We built some standard measures for AI training, which became MLPerf training.”

This collaborative spirit is echoed by Meta, a founding member of MLCommons. Meta’s commitment to benchmarking, particularly in the Chakra work group focused on improving communications performance, demonstrates the integral role of network engineering in optimizing AI performance, Gangidi concluded.

“I think benchmarking and the work that David and MLCommons team is doing is important because with benchmarks you can understand how ML models stress infrastructure or what tasks they’re able to do,” he said. “You’re able to reproduce them and repeat them. If you cannot reproduce something, then it’s hard to improve it or to make it more reliable.Bbenchmarking is a very fundamental aspect of what helps scale these clusters.”

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:

https://www.youtube.com/watch?v=videoseries

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

Photo: SiliconANGLE