Skip to main content

Qualcomm Hexagon NPU

Vendor: Qualcomm

Category: NPU Neural Processor

Architecture: Hexagon (HTA + HVX)

Introduction

Qualcomm Hexagon NPU integrated in Snapdragon SoCs, providing efficient AI inference acceleration. Accessed through Qualcomm AI Engine Direct and SNPE frameworks, supporting INT8/INT16 quantized inference.

Specifications

ModelComputeMemoryInterfaceTDPProcess
Snapdragon 8 Elite45 TOPS (INT8)LPDDR5X (shared)Integrated SoC10W3nm
Snapdragon X Elite45 TOPS (INT8)LPDDR5X (shared)Integrated SoC23W4nm

Official Website

Visit Official Website

Driver Downloads

Linux

OS Support

WindowsLinuxmacOSAndroid
✅ (Snapdragon)✅ (Android)

Version History

VersionRelease DateDescription
QNN 2.x2024Snapdragon 8 Elite support

Performance Benchmarks

ModelTaskPerformance Metric
Snapdragon 8 EliteNPU TOPS45 TOPS (Hexagon)
Snapdragon X EliteNPU TOPS45 TOPS
Snapdragon 8 EliteStable Diffusion Mobile~10s/img

Pricing

ModelReference PriceNotes
Snapdragon 8 EliteProvided with SoCFlagship phone SoC
Snapdragon X EliteProvided with SoCWindows on ARM

Quick Installation

Android / Windows on Snapdragon

Hexagon NPU is invoked through Qualcomm AI Engine Direct.

# Install QNN SDK (download from Qualcomm Developer Network)
# Supports TFLite/ONNX model compilation and deployment

Code Examples

Python (QNN Runtime)

# Run compiled model using QNN SDK
from qnn import QnnRuntime

runtime = QnnRuntime()
model = runtime.load("model.qnn")
output = model.run(input_data)

Architecture Highlights

  • Hexagon DSP + HVX: Qualcomm scalar/vector DSP architecture, supporting AI inference and signal processing
  • HTA (Hexagon Tensor Accelerator): Dedicated tensor accelerator optimized for Transformer models
  • AI Engine: Unified AI software stack covering Hexagon + Adreno GPU + Kryo CPU

Model Compatibility

Model/FrameworkSupportNotes
QNN SDK✅ NativeBest support
TFLiteNNAPI/Hexagon backend
ONNXQNN compile
Stable DiffusionMobile-optimized version
WhisperOn-device voice

If you're evaluating alternatives, the following products may also fit your scenario: