[News] Qualcomm Targets NVIDIA, AMD with LPDDR-Based AI Chips; Thermal Durability Looms as a Challenge

The battle for AI data center dominance is heating up. Qualcomm, long focused on mobile chips, has unveiled its next-generation AI inference processors — the AI200 and AI250 — taking direct aim at NVIDIA and AMD. Notably, according to Wccftech, Qualcomm has shifted away from using HBM in its new chips, opting instead for a design built around mobile-oriented LPDDR memory. Wccftech also points out that the new chips feature up to 768 GB of LPDDR memory on the accelerator package, significantly exceeding the capacity of typical HBM solutions in the industry.

As the company’s press release notes, the Qualcomm AI200 supports up to 768 GB of LPDDR per card, offering higher memory capacity and lower cost to enable greater scalability and flexibility for AI inference. The AI250, meanwhile, debuts an innovative near-memory computing architecture that delivers greater than 10x higher effective memory bandwidth and much lower power consumption.

However, Wccftech indicates that Qualcomm’s rack-scale solutions may lag behind mainstream offerings from NVIDIA and AMD, as avoiding HBM might lead to potential reliability issues under continuous high-temperature server conditions.

Inside Qualcomm’s AI200 and AI250

Both the Qualcomm AI200 and AI250 accelerators are built on its Hexagon NPUs, customized for data center AI workloads. However, as Tom’s Hardware notes, one key detail Qualcomm has not revealed about its AI200 and AI250 rack-scale inference systems is which processors will power them.

In addition to their NPU and memory architecture, Qualcomm’s press release states that both rack solutions feature direct liquid cooling for enhanced thermal efficiency, PCIe for scale-up capability, Ethernet for scale-out connectivity, confidential computing for secure AI workloads, and a rack-level power consumption of 160 kW.

Per Qualcomm’s press release, the AI200 and AI250 accelerators are slated for commercial release in 2026 and 2027, respectively. Bloomberg reports that the first customer may be Saudi Arabia’s AI startup Humain, which plans to deploy 200 megawatts of computing capacity based on the chips beginning in 2026.

Read more

• [News] Qualcomm Launches Snapdragon 8 Elite Gen 5, May Plan Samsung Galaxy-Exclusive Version
• ASIC Set to Outpace GPU? NVIDIA’s Scale-Up and Beyond

(Photo credit: Qualcomm)