How Reducing an AI Execution Bottleneck Lowers Needed Power and Cooling | Sambanova

本站添加: 2026-05-19

[00:00:00]Hey, how's it going, man?

[00:00:01]>> I'm good. Can you tell me a little bit about what you're doing over here at SambaNova? Yeah, so SambaNova are building new chips for the AI agentic era.

[00:00:09]We do that by creating an entirely new proprietary silicon that's unlike any other architecture in the world. We call it the RDU, the reconfigurable data flow unit, and the main value proposition of it is that you're actually able to run your AI inference faster than you've ever been able to before while also consuming way less energy. And this will, I assume, reflect in lower thermals, as well.

[00:00:32]Yeah, exactly. So, you're able to kind of like deploy this in like a lot of like interesting and like more remote environments, like places that data centers haven't been able to go historically. And so, we're really hopeful that this is really going to open up a lot of like advantages and allow people to continue building like more and more complex applications while, hopefully, you know, saving the environment a little bit, too. Can you explain a little bit about how the RDU accomplishes these low power? Yeah, so basically, it runs on a data flow-based execution model.

[00:01:00]And also, we have a three-tier memory structure, which allows us to switch memory in and out of access. And the combination of those two things allows us to map the entire decoder phase of an LLM onto our system compared to a GPU, which has to execute kernel by kernel.

[00:01:13]Because we've eliminated that bottleneck and overhead, that's one of the reasons our systems run so fast and efficiently.

[00:01:18]How would an engineer be able to get a hold of this to evaluate it that for themselves? Yeah, so you can learn more about us at sambanova.ai. We also have a cloud, which you can also visit from our website itself, that allows you to get an API key and start playing around with it right on your local environment.

[00:01:33]Thank you very much for showing us that.

[00:01:35]Of course, man. Hey, it's a pleasure.

[00:01:36]Thanks so much.