
Qualcomm and Cirrascale open Dragonfly AI200 for selected customer evaluations
The AMW Read
Selected customer access advances Qualcomm's inference hardware offering, but the announcement does not establish adoption or comparative performance.
Qualcomm and Cirrascale open Dragonfly AI200 for selected customer evaluations
Qualcomm and private cloud provider Cirrascale have made the Dragonfly AI200 inference accelerator available to selected customers for evaluation in a cloud environment. Qualcomm announced the access on November 17. Customers can use it to explore how their AI workloads perform on the system. The source describes AI200 as Qualcomm's first rack-scale AI solution.
The system adds a Qualcomm offering to the market for AI inference infrastructure. Each accelerator card has 768 GB of LPDDR5X memory. A liquid-cooled ORv3 rack contains 56 cards, with a reported 43 TB of total memory and 0.414 PB/s of memory bandwidth. It supports PCIe Gen6 scale-up and RoCE scale-out. Qualcomm says a single rack can run inference on models with around 10 trillion parameters and offer a lower total cost of ownership. The source does not provide comparative cost or performance results.
For builders, the Cirrascale evaluation offers a way to test their workloads against the system's memory capacity and interconnect design. Investors can watch whether customer tests establish performance and cost advantages beyond the specifications. Access is limited to selected customers, and the announcement does not establish broad adoption or production deployment.


