CREDIT: NVIDIA
TECH & AI
Validating next-generation rack architecture The practical output of these advanced digital twins can be seen in the new reference designs required to house the latest silicon. A prime example is the newly unveiled reference design for the NVIDIA Vera Rubin NVL72 racks.
Validated entirely through ETAP models for electrical system design and EcoStruxure IT Design computational fluid dynamics( CFD) for layout and airflow, the design provides a roadmap for powering and cooling ultra-dense rack-scale systems. The digital validation addresses several stringent infrastructure requirements, including:
• Elevated power and cooling: Facilitating power distribution with an increased supply voltage of 480 VAC, alongside allowing a higher TCS loop supply temperature of 45 ° C to enhance overall cooling efficiency.
• Clustered architecture: Supporting an IT room layout where clusters of AI racks share centralised networking, storage and CPU support. This keeps NVIDIA rack-scale systems physically close while allowing separate, higher voltages for the GPU racks to optimise power delivery.
• Performance optimisation: Designing facilities to accommodate various operating points of GPU racks( both MaxP and MaxQ). Operating at MaxQ, for instance, overrides power constraints to achieve more tokens per watt, optimising computing performance through built-in redundancy.
Because these configurations have already been verified in a digital twin environment, data centre owners can scale dense architectures rapidly without compromising system stability.
datacentremagazine. com 71