"fp4 precision"

#13457738

16 Mar 2026

A detailed view of the NVIDIA GB300 NVL72 liquid-cooled rack-scale system is displayed at the NVIDIA pavilion during the Mobile World Congress in Barcelona, Spain, on March 5, 2026. This computing unit integrates 72 NVIDIA GB300 Tensor Core GPUs and 36 NVIDIA Grace CPUs, interconnected via the fifth-generation NVLink interconnect to function as a single GPU. The system is powered by NVIDIA BlueField-3 DPUs and the ConnectX-8 SuperNIC, capable of delivering up to 800Gb/s of networking throughput. The rack features a direct-to-chip liquid cooling system designed to manage the thermal output of the 120kW power density. This configuration is optimized for training Large Language Models and real-time inference on trillion-parameter models, utilizing the Transformer Engine and support for the FP4 precision format. The display highlights the NVIDIA Quantum-3 InfiniBand switches at the top of the rack, facilitating ultra-low latency communication between multi-rack clusters.

Restricted to Editorial Use Only.
Commercial use is not permitted without prior authorization.
Please contact us for more information.