Qualcomm Applied sciences, Inc. has introduced the launch of its newest AI-optimised options for information centres: the AI200 and AI250 accelerator playing cards and racks. Constructing on its NPU expertise management, these choices carry unrivalled rack-scale efficiency and reminiscence capability, setting a brand new normal for generative AI inference.
The Qualcomm AI200 is tailor-made to ship low whole value of possession and optimised efficiency for big language fashions and different AI workloads. With help for 768 GB of LPDDR per card, it addresses excessive reminiscence wants with an emphasis on scalability and adaptability for AI duties.
In the meantime, the Qualcomm AI250 introduces a pioneering reminiscence structure based mostly on near-memory computing. This delivers a greater than 10x enchancment in reminiscence bandwidth effectivity and considerably reduces energy consumption, facilitating the strong software of AI in {hardware} tailor-made to fulfill various buyer wants.
Each options are outfitted with options like direct liquid cooling, PCIe for scalability, and Ethernet for enlargement. Additionally they incorporate confidential computing to safeguard AI processes, with an influence draw of 160 kW per rack. This enables Qualcomm’s choices to fulfill the calls for of contemporary information centre environments.
In keeping with Durga Malladi, SVP & GM at Qualcomm, “Our wealthy software program stack and open ecosystem help make it simpler than ever for builders and enterprises to combine, handle, and scale already skilled AI fashions on our optimised AI inference options. With seamless compatibility for main AI frameworks and one-click mannequin deployment, Qualcomm AI200 and AI250 are designed for frictionless adoption and speedy innovation.”
The mixed software program stack helps a wide selection of machine studying frameworks and inference engines, optimising for generative AI duties with methods like disaggregated serving. Builders profit from streamlined mannequin integration and entry to AI purposes, libraries, and instruments, making certain operational effectivity.
Anticipated to be commercially obtainable from 2026 and 2027, the AI200 and AI250 embody Qualcomm’s dedication to annual progress in information centre AI, specializing in efficiency, vitality effectivity, and cost-effectiveness.
