The new Hexagon NPU of the Snapdragon 8 Elite Gen 6 Series is prepared for the era of AI agents

The new Qualcomm processors for mobile phones will soon be available, the Snapdragon 8 Elite Gen 6 Series. We speak in plural because the company has decided to segment this high-end into two references, with one even more powerful if possible, the Snapdragon 8 Elite Gen 6 Pro. As a preview, the company has shown us the new features that one of the NPUs of these processorsalthough it has not specified which model it corresponds to.

Qualcomm’s Hexagon NPU prepares for the new era of AI agents

With the arrival of the AI agentsthis NPU has changed to add improvements in this aspect. The Qualcomm Hexagon NPU that will arrive with the next generation of processors, now has an element accelerator focused on transformers and a 50% more shared memory to keep model data close to the NPU. This way agents can continue working while driving more complicated contexts, more tools and even perform tasks concurrently.

It is designed to cover more AI tasks and adapt to the new generation of this technology. Element Accelerator can handle transformer workloads for generative AI and the latest agents. All this helped by vector, scalar and matrix extensions to adopt larger models and help agents provide faster responses and reason more efficiently. This will offer new, more complete user experiences on mobile devices without affecting their energy efficiency.

Qualcomm improves Hexagon NPU to accelerate larger AI models

The new Qualcomm Hexagon NPU It is also faster, with up to 50% more prefill on models that use INT 4 precision. It is also ideal for models that use MoE (Mixture-of-Experts)and that only activate a part of their parameters per token. It also offers better performance in Continuous AI, extensive reasoning, multimodal models and concurrent agents.

In conjunction with the Qualcomm CPUthe NPU Hexagon will offer a Agent AI locally, efficiently and with superior responsivenesswith innovation in tasks such as acceleration, memory management, precision and model loading.