Post by SiliconANGLE & theCUBE

Previewing CoreWeave’s first production Vera Rubin NVL72 rack ✨ During #theCUBE’s coverage from “Scaling the Agentic Era,” we hear from CoreWeave's Jacob Yundt, Senior Director of Compute Architecture, Zach Merendino, Staff Operations Engineer, Fleet Engineering, & Harshdeep Banwait, Director of Product, about the architecture of the first production Vera Rubin rack: 72 GPUs across 18 compute trays with NVLink 6 switches, managed by Racky and Valvey, CoreWeave's custom liquid cooling system. “The first thing we're looking at right up here is you have your first 10 Nvidia Vera Rubin compute trace. Each one has four Rubin GPUs and two Nvidia Vera CPUs in it. Right in the middle, you have your Nvidia NVLink 6 and NV link switches. And then, right underneath you have your eight final compute trace, and that makes up all 72 GPUs interact,” Merendino explains. “In addition to the GPU servers and the MV link switches, we also have some supporting infrastructure here. Up top, we've got the power shelves, so these are much beefier than our previous generation power shelves. One of these is now 110 kilowatts. Beneath that we've got Racky, our custom rack manager. From the leftmost side, we've got all of our sensor ports. This is for leak grope and some other things like condensation sensors,” Yundt adds. “We've got our two isolation valves here, that's for fast on/off in case something is starting to leak. Some extra I/O is here for E1.S storage for the Racky OS, and you'll notice at the bottom of the rack, previously this is where we would have the inner rack CPU, but now this is all cooled by Valvey. So, Racky is connected to Valvey. Valvey is responsible for modulating the flow and temperature, making sure that this GPU rack is configured optimally and running cool,” he furthers. 💡 Get more insights! https://lnkd.in/erWUyyEB #VeraRubin #Racky #Valvey #CoolingSystem #EnterpriseAI

Video Content