Qualcomm unveils Hexagon NPU to power mobile Agentic AI

Qualcomm has unveiled details regarding its next-generation Hexagon NPU (Neural Processing Unit), designed to enable “Agentic AI” capabilities on mobile devices. This new architecture focuses on allowing AI assistants to understand user context, reason across multiple applications, and execute complex tasks autonomously.

A key technical advancement is the introduction of the Element Accelerator, which is specifically optimized for Transformer models—the foundation of generative AI. This unit works alongside scalar, vector, and matrix extensions to accelerate large language model (LLM) workloads while maintaining energy efficiency. Additionally, the NPU features a 50% increase in shared memory capacity, allowing more data to remain on the processor to reduce latency and improve context understanding.

The new NPU also supports the Mixture-of-Experts (MoE) architecture. This allows the system to run models with up to 30 billion parameters by only activating approximately 3 billion parameters per generated token, significantly reducing the computational burden. For INT4 models, Qualcomm reports a 50% improvement in prefill performance, enabling AI responses to begin within 1.5 seconds.

Further details regarding the new Hexagon NPU and the Snapdragon 8 Elite series are expected to be announced at the Snapdragon Summit, scheduled for September 22–24.