Qualcomm continues to gradually demonstrate the capabilities of its next-generation Snapdragon processors for high-end smartphones. The official demo is expected to take place during the Snapdragon Summit on September 22there are two Elite variants, which can be called Snapdragon 8 Elite Gen 6 and 8 Elite Gen 6 Pro.
After first anticipating CPUs capable of exceeding 5 GHz and GPUs with new frame-scaling technology, Qualcomm is now focusing on units dedicated to neural processing. The goal is Add artificial intelligence capabilities that run directly on the deviceespecially for proxy-based applications.
NPU designed for more complex artificial intelligence models
The new Hexagon NPU introduces an element called Element Accelerator, which complements the scalar, vector and tensor units already in the architecture. Its job is to speed up the inference of the Transformer model, Reduce the time required to process AI proxy requests According to Qualcomm, there will be no increase in energy consumption.
Another change involves memory shared between the four processing units. More memory lets you Reduces the need for continuous access to RAM Smartphone performance, especially when managing KV cache. In practice, NPUs can retain more information about the context of an ongoing conversation or activity, which is an important feature for models that can manage multi-step operations.
Snapdragon lags behind MediaTek, Samsung grows with Exynos: But the smartphone chip market is in crisis
Qualcomm also designed the new architecture to work with Mix of Experts (MoE) models. This technique does not activate all parameters of the model simultaneously, but Route each request to the most appropriate “expert.”. Models with up to 30 billion parameters can be run locally with less computational and memory overhead than running the model entirely.
In terms of performance, Qualcomm claims 50% faster The amount of prefill, greater decoding throughput, faster speculative decoding, and an increase in the total number of tokens processed per second. However, the 50% figure specifically relates to INT4-accurate models and therefore does not equate to a general 50% increase in AI performance.
The overarching goal of this intervention is to Make AI functions that can be executed without connecting to the cloud more responsive. For Qualcomm, the next step is not just to respond to user requests faster, but to enable smartphones to run artificial intelligence agents that can understand context, plan multiple actions and execute them locally.
The information about the new NPU is only a preview for now: Qualcomm should provide all the details of the next generation Snapdragon at the Snapdragon Summit on September 22, as well as the full characteristics of the two variants of the new generation.
