UK — The Register
Qualcomm offers test drive of its rack-scale Nvidia AI challenger
Chips1 min read
Qualcomm has opened cloud trials of its Dragonfly AI200 server rack to selected customers. The goal is to run AI models with less power than Nvidia's systems.
What happened
Qualcomm let chosen customers test its Dragonfly AI200 rack-scale system over the cloud. The test runs on Cirrascale's AI Innovation Cloud. Cirrascale also showed the system at the SC26 computer conference in Chicago on November 17 and 18. Qualcomm says customers who like the results may move on to full production use.
Key numbers and facts
- Each AI200 accelerator card (a chip board that speeds up AI work) has 768 GB of LPDDR memory.
- 56 cards together give about 43 TB of memory in one liquid-cooled rack.
- Qualcomm says the system can run AI models from 7 billion to 10 trillion parameters (the numbers that set how a model learns).
- The Dragonfly line also includes the AI250 and AI300 accelerators and the C1000 server CPU.
Background
AI inference means using a trained AI model to answer questions or make outputs. It is becoming one of the biggest jobs in data centers, the huge buildings full of computers. Nvidia leads this market. Its Blackwell-based NVL72 systems use fast but expensive HBM3e memory. Qualcomm chose LPDDR memory instead, which is more common in phones. This keeps costs and power use lower, but it does not automatically match Nvidia's speed.
Why it matters
AI systems use huge amounts of electricity, and running costs can be very high. If Qualcomm's approach saves power and money, more companies may pick it instead of Nvidia. The test will show whether the promised savings are real.
What to watch
- Whether test customers report real power and cost savings.
- Whether the AI200 can match Nvidia's speed on real AI tasks.
- Whether the trials turn into full production deployments.
- How the Dragonfly AI Infrastructure Management Suite software performs in large setups.
Glossary
- Qualcomm
- A chip company that makes processors for phones and now also for data centers.
- Cirrascale
- A cloud company that rents AI computing power to customers and offers test access to new systems.
- Nvidia
- A chip company and the leader in making chips used for AI.
- LPDDR
- A type of low-power memory that is more common in mobile phones.
- HBM3e
- A type of fast but costly memory stacked close to the chip, used in high-end AI systems.
- Dragonfly AI Infrastructure Management Suite
- Qualcomm's software for setting up, watching, and managing AI racks.