IT之家 · 智能时代T2·Infra
58 pts
Qualcomm, together with AI private-cloud provider Cirrascale, has opened cloud-based evaluation of its first rack-scale inference solution, Dragonfly AI200, to select customers. Each card carries 768GB of LPDDR5X memory, and a 140kW liquid-cooled rack packs 56 cards with 43TB total memory and 0.414 PB/s bandwidth, targeting 10T-parameter-class inference in a single rack at lower total cost of ownership.