Boost performance with the Qualcomm Hexagon NPU Driver 1.0.0.12 on Windows PCs powered Snapdragon X Series

Co-written with Sandheep Balasubramanian and Anurag Nikhil Darbha.

 

Qualcomm Technologies is making rapid strides in performance improvements with the Qualcomm Hexagon NPU Driver 1.0.0.12 Public Release package!

With improvements to latency, throughput and power efficiency, the latest driver provides significant competitive advantages to running AI workloads on devices powered by Snapdragon X Series, powered by the Hexagon NPU accelerator.

Release version: 1.0.0.12

Platforms: Snapdragon X Series

Operating System: Windows 11 24H2

Download the NPU driver from Qualcomm Software Center.

Follow the installation instructions provided in the Release Notes.

AI performance where it matters most to users

AI workloads at the edge rely heavily on low latency, high throughput and low power utilization for the best user experience. These metrics are ubiquitous standards in how PC users measure performance of their applications.

The latest Hexagon NPU driver takes these challenges head-on and provides significant performance improvements release-over-release. Let’s explore these key features further!

Key features:

• A shared memory optimized digital signal processor (DSP) task queuing-based IPC path in Qualcomm Neural Processing SDK (QNN)/Microsoft Compute Driver Model (MCDM)  stack for Host to NPU messaging. This reduces RPC overhead and improves latency
  
  The queue is automatically enabled where supported with a seamless fallback to legacy FastRPC based messaging in scenarios where the queue may not be available or times out
  
  
• Proactive memory trimming for handling pressure scenarios efficiently through DirectX-Kernel Trim support
  
  
• Introduction of a low-priority execution mode for bursty or background tasks, optimizing resource usage
  
  
• Enhanced scalability with more than 10 simultaneous NPU processes available for high-throughput workloads. This allows seamless performance with expanded support for multiple concurrent applications enabling effortless multi-session support for a smoother, uninterrupted experience

Impact on latency, throughput, power efficiency, source-code compatibility:

• Latency: A significant reduction in IPC transport latency is observed when the optimized DSP task queuing is used, contributing to measurable end‑to‑end inference improvements for some models, depending on platform, mode and workload
  
  
• Throughput: For pipelined or bursty workloads, higher steady‑state throughput may be observed due to reduced per‑transaction signaling and batch processing in callbacks
  
  
• Power and CPU: Increase CPU/DSP utilization. Default policies established by the operating system balance latency and power; power‑sensitive deployments can prefer interrupt/callback modes
  
  
• Source-code compatibility: No source‑level changes are expected for applications using standard QNN interfaces. The Hexagon NPU Driver may require a newer version of the Qualcomm AI Runtime Development Package (QAIRT) SDK to realize the features and gains mentioned above

Read the release notes (packaged with the installer download) for more details on improvements packed into this driver update.

Your feedback is valuable!

Take the driver for a spin and let us know what you think.

You can connect with the developer community through our Windows on Snapdragon support forum or join real-time conversation with global developer community at Developer Discord.