CoreWeave has announced it will be among the first providers to offer large-scale NVIDIA Blackwell clusters, featuring two configurations: the NVIDIA HGX B200 and the NVIDIA GB200 NVL72. Both configurations are interconnected with NVIDIA Quantum-2 InfiniBand networking, promising significant advancements in AI, data processing, and high-performance computing.
The NVIDIA HGX B200, designed for the most demanding workloads, leverages NVIDIA Blackwell GPUs to deliver 15 times faster real-time inference on trillion-parameter models. Meanwhile, the NVIDIA GB200 NVL72 is a liquid-cooled, rack-scaled solution that integrates 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell GPUs, achieving up to 30 times faster real-time trillion-parameter LLM inference. Introduced earlier this year at NVIDIA GTC, the Blackwell architecture can link up to 72 GPUs in a single NVIDIA NVLink domain, enabling enhanced GPU-GPU communication bandwidth. This innovation supports a wide range of applications beyond AI training and inference, including quantum computing, drug discovery, and fusion energy.
Key Points:
- CoreWeave to deliver NVIDIA Blackwell clusters with configurations HGX B200 and GB200 NVL72.
- Both configurations use NVIDIA Quantum-2 InfiniBand networking for high-performance interconnectivity.
- NVIDIA HGX B200 offers 15 times faster real-time inference for trillion-parameter models.
- NVIDIA GB200 NVL72 combines 36 Grace CPUs and 72 Blackwell GPUs for 30 times faster LLM inference.
- Blackwell architecture supports extensive GPU-GPU communication for advanced applications beyond AI.
