CoreWeave Deploys NVIDIA's New Grace Blackwell GPUs, Boosting AI Performance by 300%

CoreWeave, a data center cloud service company, has announced the deployment of thousands of NVIDIA's new Grace Blackwell (GB200) graphics processing units (GPUs), making it the first cloud service provider to offer this new GPU on a large scale. This announcement was made jointly by both companies.

Several artificial intelligence enterprises, including Cohere, ibm, and Mistral AI, have already begun adopting the nvidia GB200 NVL72 system. Cohere, a Canadian company focused on developing large language models (LLMs) for enterprises, reported that the training performance based on the GB200 GPU has tripled compared to NVIDIA's previous generation Hopper GPU.

Autumn Moulder, the engineering vice president of Cohere, commented, "This upgrade has brought significant performance improvements to our entire technology stack—whether it's running vertical integration North applications on a single Blackwell GPU or scaling up to thousands of GPUs for large-scale training tasks, efficiency has leaped forward."

Ask Aime: "Will NVIDIA's Grace Blackwell GPUs boost Cohere's AI performance?"

IBM is utilizing the Grace Blackwell rack system to train its next-generation enterprise-level Granite AI model. Sriram Raghavan, the vice president of AI at IBM Research, stated, "Our collaboration with coreweave will further enhance IBM's technological capabilities, helping to build more advanced, high-performance, and cost-effective AI models, supporting enterprise-level and agent AI applications on the IBM watsonx platform."

Mistral AI, a company based in Paris, is also among the first customers to train and deploy models on the new system. Timothee Lacroix, the co-founder and chief technology officer of Mistral AI, noted, "Without any additional optimization, we have observed a twofold increase in the performance of dense model training."

Michael Intrator, the CEO and co-founder of CoreWeave, stated, "CoreWeave is committed to accelerating technological iterations, and this first large-scale deployment of the most advanced system once again demonstrates our commitment."