CoreWeave has deployed racks packed with Nvidia Blackwell Ultra GPUs for their cloud AI service, in partnership with Dell and Switch.
Each rack is densely packed with 72 Blackwell Ultra cards containing 36 Arm-based 72-core Grace CPUs and 36 BlueField DPUs, along with liquid cooling for managing the power generation.
The new setup is expected to provide around 50% higher performance than the previous architecture, delivering 1.1 ExaFLOPS of dense FP4 inference and 0.36 ExaFLOPS of FP8 training performance per rack.
The goal is to enable faster and more effective training of large language models, enhance AI reasoning abilities, and improve scale-out capabilities with speeds reaching up to 14.4 GB/s.