Qualcomm has joined forces with Cirrascale, an AI private cloud provider, to give select customers access to evaluation of the Dragonfly AI200 inference accelerator. Through a cloud environment, these customers can test how their AI workloads perform on the chip. The Dragonfly AI200 marks Qualcomm's first rack-level AI solution, with each card featuring 768GB of LPDDR5X memory and supporting PCIe Gen6 scale-up and RoCE scale-out. A single 140kW liquid-cooled ORv3 rack houses 56 cards, delivering a total of 43TB memory and 0.414PB/s memory bandwidth—enough to run inference on 10T-level ultra-large models within a single rack at a lower total cost of ownership.
