NVIDIA's Blackwell Ultra rack with 72 GPUs and 36 Grace CPUs in one NVLink domain, 20 TB of HBM3e, 130 TB/s NVLink and 1,440 PFLOPS sparse FP4.
GB300 NVL72 connects 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs in a single liquid cooled rack as one 72 GPU NVLink domain. NVIDIA publishes 20 TB of GPU memory at up to 576 TB/s, 17 TB of LPDDR5X CPU memory, 37 TB of fast memory, 130 TB/s of NVLink bandwidth, 2,592 Arm Neoverse V2 cores, and 1,440 PFLOPS of sparse FP4 (1,080 PFLOPS dense) per rack.
The rack is designed for test time scaling inference and for training runs that benefit from a large NVLink domain. It is delivered as a complete system with compute trays, NVLink switch trays, power shelves and the coolant distribution interface, and it requires a liquid cooled data center with the matching facility water loop. We quote GB300 NVL72 as a project with site requirements and a Quantum-X800 or Spectrum-X scale out fabric, delivered in phases and commissioned in your liquid cooled hall.
How to decide between air and direct liquid cooling for GPU servers, from the rack power limit up to NVL72 racks that only come liquid cooled.
In this guide
We quote every class that carries it, from a single node to a pod. Tell us the workload and we propose the chassis.
