Article
AI News Chipmakers

Qualcomm lets customers test its Nvidia challenger in the cloud

Selected customers can trial Qualcomm's Dragonfly AI200 inference system through Cirrascale, before any pricing has been published.

by TechDefused · Edited by IL
The image displays a server room equipped with various networking equipment and servers. The focus is on a rack containing multiple devices connected by cables, indicating a well-organized technological setup. — Credit: Photo by Kevin Ache on Unsplash c Photo by Kevin Ache on Unsplash

Qualcomm has opened cloud-based trials of its Dragonfly AI200, a rack-scale system it is pitching as a cheaper alternative to Nvidia for running AI models, The Register reported.

Selected customers can test the system through Cirrascale Cloud Services, with a route to full production deployment.

Cirrascale is a cloud provider specialising in dedicated infrastructure for AI training and fine-tuning and for inference, the work of running trained models to produce answers.

It already supports accelerators from Nvidia, AMD and Tenstorrent, and is hosting the Qualcomm trials on its AI Innovation Cloud.

Qualcomm says the AI200 uses less power and costs less than Nvidia-based systems for handling large language models, multimodal models and AI agents, and for retrieval-augmented generation (RAG).

Each accelerator card carries 768 GB of LPDDR memory, the type more commonly found in mobile devices.

Nvidia relies on HBM3e, a faster and more expensive form of high-bandwidth memory.

A liquid-cooled rack of 56 cards provides about 43 TB of memory and can run models with anywhere from 7 billion to 10 trillion parameters, the internal values that determine a model's size.

"AI inference is becoming one of the largest and most important computing workloads in the datacenter, and the industry needs a more efficient approach," said Tony Pialis, Qualcomm's executive vice-president and general manager for data centres.

Qualcomm has not disclosed pricing for the AI200 or any independently verified performance figures.

The trials let customers benchmark the system on their own workloads rather than relying solely on Qualcomm's specifications.

The AI200 is part of Qualcomm's Dragonfly data centre line-up, alongside the AI250 and AI300 accelerators and the C1000 server processor.

It forms part of a wider push into custom chips for the biggest cloud providers, which includes a chip development partnership with Amazon Web Services.

Read more about Qualcomm

by TechDefused