Brief IA
›

Qualcomm Launches Dragonfly AI200 for Cloud Testing

Models & LLM·Tom Levy·

Qualcomm Launches Dragonfly AI200 for Cloud Testing

Qualcomm Launches Dragonfly AI200 for Cloud Testing
Key Takeaways
1Qualcomm is reserving cloud trials of the Dragonfly AI200 for selected clients, through Cirrascale.
2Each card has 768 GB of LPDDR; a rack holds 56 cards.
3Qualcomm announces a potential move to production after evaluation.
Why it matters : this access allows the relevant clients to assess the savings claimed by Qualcomm for inference.

Qualcomm is making its Dragonfly AI200 inference system accessible to selected clients through Cirrascale Cloud Services. The manufacturer presents its system as a less power-intensive and cost-effective alternative to Nvidia systems, savings that these trials are intended to evaluate.

Executives Discuss Inference Costs

For Dave Driggers, CEO of Cirrascale, clients should be able to leverage increasingly sophisticated models on an efficient, scalable infrastructure without being hindered by energy or operational expenses. He presents the collaboration with Qualcomm as an opportunity to assess an architecture designed for the next generation of inference.

Tony Pialis, executive vice president and general manager of data centers at Qualcomm, believes that inference is becoming a major computing activity in these facilities and requires greater efficiency. According to him, the AI200 brings the group's expertise in high-performance, low-power computing to the rack. He describes an architecture designed to meet the performance, memory, and cost needs of inference.

Qualcomm Focuses on 768 GB per Card

Aimed particularly at large language models and multimodal models, the AI200 features 768 GB of LPDDR per card. A liquid-cooled rack houses 56 cards, providing a capacity of approximately 43 TB. Qualcomm claims compatibility with models ranging from 7 billion to 10 trillion parameters, as well as with long contexts, retrieval-augmented generation (RAG), and processes involving agents.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

However, this memory capacity does not guarantee matching Nvidia in performance. The competitor's NVL72, based on Blackwell, utilizes fast but expensive HBM3e stacks. The LPDDR chosen by Qualcomm is more commonly used in mobile devices.

From Trials to Production and Management Tools

The evaluation takes place through Cirrascale's AI Innovation Cloud: according to Qualcomm, participants can measure the promised savings and have the option to move to production. Cirrascale also showcased the system in Chicago during SC26 on November 17 and 18.

The AI200 belongs to the Dragonfly family intended for data centers, alongside the AI250 and AI300 accelerators and the C1000 server processor. Qualcomm detailed its broader ambitions in June during its investor day. The group also offers the Dragonfly AI Infrastructure Management Suite, a software suite that ensures provisioning, monitoring, orchestration of processes, and failure management in rack installations.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.