Qualcomm intends to unveil its next-generation flagship chips at the Snapdragon Summit on September 22.
According to ETNA via Android Authority, although the exact names of these chips are not yet known, existing evidence indicates one of the main focuses of the new generation will be on on-device AI enhancements and running AI agents without a persistent dependency on cloud servers.
The heart of these chips' AI capabilities is a new generation Hexagon NPU designed to run multifaceted AI applications and especially AI Agents. One of the most important changes is the addition of an Element Accelerator alongside Scalar, Vector, and Tensor units. Qualcomm says this accelerator can speed up inference of Transformer-based models and allow agents to deliver faster responses and more efficient processing without increasing power consumption.
Another significant change is a substantial increase in the NPU's shared memory. This architecture reduces reliance on the phone's main memory when exchanging KV-cache related data and allows the NPU to hold more context. The result of this change could be the ability to run heavier workloads, particularly multi-step tasks of AI agents.
Qualcomm has also introduced support for new Mixture-of-Experts (MoE) architectures. In this approach, not all parameters of a model are activated simultaneously and only the portion needed for each request is used. According to Qualcomm, this architecture can enable running models with up to 30 billion parameters locally on a phone.
The new generation NPU also delivers up to 50 percent higher preprocessing performance, faster decoding, improved speculative decoding, and more tokens per second in models using INT4 precision. Of course, this figure does not mean a direct 50 percent increase in performance across all AI capabilities.
Overall, Qualcomm's goal is for future flagship phones to be able to run not only simple AI queries but also multi-step and agent-driven processes faster and more locally.