Qualcomm Unveils 32K-Context Hexagon NPU for Snapdragon AI as 5GHz Mobile CPU Nears
Updated
Updated · Computerworld · Sep 10
Qualcomm Unveils 32K-Context Hexagon NPU for Snapdragon AI as 5GHz Mobile CPU Nears
3 articles · Updated · Computerworld · Sep 10
Summary
Qualcomm said its next Hexagon NPU adds a new Element Accelerator for transformer models, targeting fast action loops, multimodal workloads and context lengths of up to 32K on flagship Snapdragon phones.
50% more shared memory and up to 50% faster INT4 pre-fill versus Snapdragon 8 Elite Gen 5 are meant to keep KV-cache and model state closer to the NPU, cutting latency and power use.
Qualcomm is also aiming at Mixture-of-Experts models up to 30 billion parameters, with about 3 billion active per token in its example to make larger on-device AI workloads more practical.
The NPU reveal completes Qualcomm's staggered preview of its next premium mobile platform, following a 5GHz Oryon CPU and an Adreno GPU revamp with AI matrix cores and claimed power savings of up to 40%.
The broader pitch is a tighter CPU-GPU-NPU design for agentic AI that could help Android partners such as Samsung and Xiaomi compete more effectively with Apple when the full chip debuts later this month.