高通发布新一代 Hexagon NPU（Element Accelerator、共享内存+50%、首次支持 30B 参数 MoE），细节留待 9 月 22-24 日骁龙峰会

Qualcomm unveiled its next-generation Hexagon neural processing unit (NPU) on Sept. 10, designed to keep agentic artificial intelligence (AI) running continuously on mobile devices.
Qualcomm Hexagon is a mobile NPU used in Snapdragon platforms. The new Hexagon introduces an Element Accelerator, 50% more shared memory and support for Mixture-of-Experts (MoE) architectures for the first time.
The key focus is optimizing agentic AI that understands a user's situation and context, reasons across multiple applications and performs tasks on the user's behalf.
Adds AI Accelerator and Expands Shared Memory
The Element Accelerator is specialized for neural-network transformer model workloads. Transformers are neural-network algorithms that process relationships among words and data in context and form the backbone of generative and agentic AI.
The Element Accelerator works alongside Hexagon's scalar, vector and matrix extensions. It accelerates large language model (LLM) computation while helping maintain power efficiency on mobile devices.
Hexagon's shared memory capacity has increased by 50%, allowing more information to remain within the processor. This reduces the need to move AI model state, activations and intermediate tensor data to and from external DRAM. The design minimizes transfer latency and data bottlenecks, allowing AI to retain longer conversational context while responding and switching tasks more quickly.
MoE Architecture Supports Models With Up to 30B Parameters
The type of AI models processed by the NPU has also changed. Previous systems used dense models in which a single large language model handled all tasks. The new Hexagon connects multiple specialized AI models suited to specific tasks and contexts using an MoE architecture.
The new Hexagon can run MoE models with as many as 30 billion, or 30B, parameters. For each token generated, only about 3 billion, or 3B, routed parameters are activated for computation.
The architecture also combines intelligent NAND flash-to-memory management and caching. Only the specialized AI models needed for a task are loaded from NAND flash into DRAM in real time, while frequently used sub-models, or expert parameters, remain in cache memory.
This reduces power consumption and allows large-model-class AI performance with lower memory bandwidth. The technology was developed in collaboration with major memory manufacturers and AI model developers.
New Architecture Broadens Data-Precision Support
Support for data precision has also been expanded. The NPU supports 16-bit floating point (FP16) for high-precision computation, as well as 2-bit integer (INT2), INT4, INT8 and FP8 formats.
For INT4-based models, prefill performance, which covers reading a prompt and performing initial computation, is up to 50% faster. Decoding throughput for generating response text and speculative decoding, which predicts future tokens for parallel processing, have also improved. Responses can begin within 1.5 seconds.
The new Hexagon will also work more closely with other compute units in the Snapdragon platform. The central processing unit (CPU), based on Qualcomm's Oryon core architecture, orchestrates multistep AI workflows and keeps data ready for the NPU. The Adreno graphics processing unit (GPU) will adopt dedicated acceleration based on NPU architectural concepts. This enables Adreno Neural Fusion, which directly integrates AI into mobile-game graphics processing.
"With the new Hexagon NPU, we have laid the foundation for agentic AI," said Vinesh Sukumar, vice president of product management for AI at Qualcomm Technologies. "With our next-generation premium mobile platform, Snapdragon, we will enable agentic AI experiences directly on smartphones." Further details will be unveiled at Qualcomm's Snapdragon Summit in the US from Sept. 22 to 24.