Phone chips are no longer just about clock speeds and core counts. Qualcomm knows this, and ahead of its Snapdragon Summit later this month, it’s making that point loud and clear by detailing the Hexagon NPU inside the upcoming Snapdragon 8 Elite Gen 6 Pro. The headline numbers are significant: 50% more shared memory, support for Mixture-of-Experts models up to 30 billion parameters, and prefill speeds that are reportedly up to 50% faster on certain models.
As reported by Android Headlines, the biggest structural addition to the Hexagon NPU is something Qualcomm calls the Element Accelerator. It sits alongside the existing scalar, vector, and tensor processing units. Its specific job is accelerating transformer models, which are the architecture behind chatbots, AI agents, and most of the on-device AI experiences phone makers are pushing right now. This is a targeted architectural choice, not a general performance bump, and that distinction matters.
The 50% increase in shared memory is also worth paying attention to. More on-chip memory means the NPU can keep AI model data closer to where computation happens, rather than pulling from the phone’s main RAM. That reduces latency. For agentic AI use cases, where a model needs to respond quickly to multi-step instructions without a cloud connection, lower latency is the entire ballgame.
To put this in context, Apple’s A18 Pro and MediaTek’s Dimensity 9400 are already shipping with NPUs tuned for on-device model inference. Google’s Tensor G5, built with Samsung, is also pushing hard in this direction. Qualcomm has been competitive, but the jump to 30 billion parameter MoE support on a phone chip is a meaningful claim. Most current flagship devices struggle to run models above 7 to 13 billion parameters smoothly on-device. If the Gen 6 Pro delivers on that number, it changes what developers can build for Android without relying on server-side inference.
The full chip is expected to be officially announced at the Snapdragon Summit from September 22 to 24 in Maui. Qualcomm is reportedly splitting its flagship line into two chips this cycle:
- Snapdragon 8 Elite Gen 6 Pro: Adreno 850 GPU, LPDDR6 memory, peak clock speeds rumored near 5GHz, 2nm TSMC process
- Snapdragon 8 Elite Gen 6: Adreno 845 GPU, LPDDR5X memory, also on 2nm TSMC
So the NPU details Qualcomm is sharing now are specifically for the Pro variant, and that tiering makes sense. The developers and OEMs most likely to ship aggressive on-device AI features will reach for the top chip. Qualcomm is giving them a clear reason to do so. Whether the real-world performance matches the spec sheet is a question that benchmarks will answer. But the architectural direction is clear, and it’s pointed squarely at making your phone a capable AI compute device without a data center behind it.
