Overview
- Qualcomm confirmed a September 22 reveal at its Snapdragon Summit and over recent days publicly teased technical details for the Hexagon NPU that will ship in the Snapdragon 8 Elite Gen 6 family, including a Pro SKU.
- The company highlighted a new Element Accelerator that it says is purpose‑built to speed up transformer workloads, the class of neural networks that power chatbots and agent‑style AI.
- Qualcomm claims the NPU has roughly 50% larger shared memory to keep frequently used model data near the processor and up to 50% faster INT4 prefill speeds to speed low‑precision model start‑up.
- The firm says the NPU will support Mixture‑of‑Experts (MoE) models of roughly 30 billion parameters by activating only portions of a model per token, a technique that reduces compute and memory needs for large models.
- Independent benchmarks, thermal and power tradeoffs, and full specs remain pending hands‑on testing after the September 22 reveal, but the changes reflect a broader industry push to move more multimodal AI workloads off the cloud and onto phones.