Qualcomm upgrades Snapdragon NPU for faster local AI agents

Qualcomm is preparing to unveil its next-generation mobile processors at its annual Snapdragon Summit on September 22. Ahead of the official launch, the company shared details about specific improvements to the Hexagon Neural Processing Unit within the Snapdragon 8 Elite Gen 6 Pro. These changes are designed to handle the increasing demands of artificial intelligence that runs directly on the phone, rather than relying solely on cloud servers.

The primary goal of these upgrades is to support what Qualcomm calls agentic AI. This refers to AI systems that can perform complex tasks, reason through problems, and manage multiple actions simultaneously. By enhancing the hardware that processes these tasks, Qualcomm aims to provide a more responsive user experience without draining the battery or overheating the device.

Hardware designed for complex reasoning

The new NPU features a component called the Element Accelerator. This part of the chip is specifically built to handle transformer workloads, which are the core calculations behind modern generative AI. It works alongside other processing units to speed up the operations that large language models rely on for understanding and generating text.

A key improvement is a 50% increase in shared memory. This extra storage space keeps frequently used AI data close to the processor, reducing the time it takes to access information. For users, this means that AI agents can manage longer conversations and multiple tasks at once without slowing down. The catch is that this hardware focus benefits complex, local tasks, but it does not change the fundamental limits of what a mobile chip can achieve compared to powerful data centers.

Efficiency gains for mobile users

Qualcomm also highlighted performance boosts for specific types of AI models. The new NPU offers up to 50% faster prefill speeds for models that use 4-bit integer precision. This technique reduces the amount of data the chip needs to process, making it more energy-efficient. It is also optimized for Mixture-of-Experts models, which activate only a fraction of their total parameters for each token generated. This design choice helps maintain speed while managing power consumption.

The company states that the NPU is designed for always-running AI, long-context reasoning, and concurrent agents. This suggests a future where phones can handle background AI tasks seamlessly. However, the trade-off remains that these advanced features will likely be reserved for premium devices. As GSMArena notes, while the technology is impressive, the availability and real-world impact on average battery life will need to be tested once the chips are widely adopted in smartphones.

Preparing for on-device intelligence

These updates signal a shift in how mobile processors are engineered. The focus is moving from raw speed to sustained efficiency for AI workloads. By integrating specialized accelerators and larger memory buffers, Qualcomm is addressing the bottleneck of keeping AI responsive on the go. This approach aims to deliver richer experiences without compromising the power efficiency that is critical for mobile devices.