NVIDIA contributes its liquid-cooled GB200 NVL72 rack design to the Open Compute Project

2024-10-15 · Publisher : NVIDIA · Domain : Cloud & DevOps

At the 2024 OCP Global Summit, NVIDIA contributes design elements of its GB200 NVL72 rack, including liquid-cooling specifications, to the Open Compute Project and extends Spectrum-X networking to OCP standards.

### Background On 15 October 2024, at the OCP Global Summit in San Jose, NVIDIA announces it is contributing foundational elements of its Blackwell platform to the **Open Compute Project (OCP)**. It shares key parts of the GB200 NVL72 electro-mechanical design: rack architecture, compute and switch tray mechanicals, liquid-cooling and thermal environment specifications, NVLink cable cartridge volumetrics. NVIDIA also extends its Spectrum-X networking platform alignment with the OCP SAI and SONiC standards. ### Key facts - System: GB200 NVL72, a liquid-cooled rack linking 36 Grace CPUs and 72 Blackwell GPUs. - Contribution: rack architecture, trays, liquid-cooling and thermal environment specifications. - Networking: Spectrum-X supports SAI and SONiC; ConnectX-8 SuperNIC cards reach up to 800 Gb/s. - Precedent: NVIDIA had already contributed its HGX H100 baseboard specification to the OCP. - Setting: the OCP Global Summit ran from 15 to 17 October 2024. ### Why it matters Liquid-cooled AI racks force a rethink of power, cooling and room layout. By publishing its specifications, NVIDIA lets other manufacturers and data centre operators design compatible infrastructure without starting from scratch. For IT departments and services firms, it is a standardization signal: future AI data centres will rely on shared templates. ### Takeaway The hardware building block of AI, the liquid rack, enters an open and shared design framework.

Tags : Datacenter, Refroidissement liquide, NVIDIA, Open Source, Puces