Page-1

# Snapdragon X2 AI/ML software solutions Provides an overview of the Snapdragon X2 AI software stack and the Qualcomm tools for accelerating, quantizing, and optimizing AI/ML models. ## Snapdragon X2 AI software stack architecture The Snapdragon X2 AI software stack routes both general AI and audio or always-on camera workloads from ML frameworks, through the Qualcomm AI Runtime, down to the CPU, GPU, NPU, and LPASS AI accelerator hardware. The stack delegates each workload to the backend best suited for it and keeps a common API and SDK across backends. The following figure shows the Snapdragon X2 AI software stack from the ML frameworks layer down to the hardware backend. **Figure:: AI software stack** At the top, general AI and audio or always-on camera use cases feed ML frameworks such as TensorFlow, Caffe, and ONNX. The API/SDK layer is the Qualcomm^®^ AI Runtime SDK. The runtime delegation layer splits work into the Qualcomm AI Runtime CPU, GPU, and Hexagon Tensor Processor paths, plus the Qualcomm AI Runtime internals, and provides CPU, GPU, and NPU operators. The implementation layer provides `QML`, `Kernels`, `NEON`, `OpenCL`, and the Qualcomm AI Runtime internal implementation. The hardware backend layer runs on the CPU, GPU, NPU, and LPAI (eNPU). ## Qualcomm AI Runtime 2.X Qualcomm AI Runtime (QAIRT) 2.X is designed for modularity, scalability, extensibility, and portability. Use it to accelerate AI/ML across the CPU, GPU, and Hexagon Tensor Processor on Snapdragon X2 and to build custom operators for the Hexagon Tensor Processor. Features: - Supports Snapdragon X2 and other devices in multiple tiers. - Provides hardware acceleration of AI/ML on Snapdragon X2 through the Nuvia CPU (FP32), Adreno GPU (FP16/FP32), and Hexagon Tensor Processor (INT4/INT8/INT16 (16A8W)/FP16). - Provides full support for user-defined operators for custom operator design and implementation on the Hexagon Tensor Processor. - Supports mixed precision activations (INT8, INT16 (16A8W), and FP16) on the Hexagon Tensor Processor. Enablement: - Snapdragon X2 requires Qualcomm AI Runtime (QAIRT) version 2.X. Documents and resources: - Qualcomm AI Runtime 2.X is available for download from Qualcomm Package Manager. ## Qualcomm AI Model Efficiency Toolkit Pro Qualcomm AI Model Efficiency Toolkit (AIMET) Pro is the closed-source, proprietary version of the public AIMET from QuIC ([https://github.com/quic/aimet](https://github.com/quic/aimet)) that includes additional benefits and early access to new features. AIMET Pro provides APIs for compression, quantization, and visualization of PyTorch and TensorFlow-trained models, so you can fine-tune models for use with Qualcomm AI software SDKs. Key features cover three areas: - **Post-training quantization**: Provides post-training algorithms such as cross-layer equalization, bias correction, adaptive rounding (Beta version), and automatic mixed precision (among INT8, INT16 or INT8, FP16) for models to be used after the quantization process through the Neural Processing Engine and AI Engine direct. - **Quantization aware training**: Trains and fine-tunes models by quantization noise simulation. It’s also available with range learning, where the quantization parameters are learned with the weights during model fine-tuning for robustness and performance improvement. - **Compression**: Provides spatial singular value decomposition (SVD), a tensor decomposition technique that splits a large layer into two smaller ones, and channel pruning, which removes redundant input channels from a layer and reconstructs weights with minimal accuracy loss. For more information about AIMET Pro, see the Qualcomm AI Model Efficiency Toolkit PRO Feature Overview (80-PK882-19). Last Published: Aug 27, 2026 [Previous Topic AI/ML software introduction](https://docs.qualcomm.com/bundle/publicresource/80-A0399-65/topics/introduction.md)