UK What Qualcomm Hexagon is
Qualcomm Hexagon is the family of processor cores Qualcomm builds into its chips for signal processing and AI inference. It first shipped in Snapdragon phone processors as a digital signal processor, handling audio, modem, camera and sensor work at very low power. As neural networks arrived on phones, Qualcomm added vector and then tensor hardware to it, and it now goes by the name Hexagon NPU.
Qualcomm’s processors were designed for phones, where battery life comes first, so AI work goes to Hexagon rather than the CPU. Its vector and tensor hardware runs quantized models at low power. Developers reach it through Qualcomm’s AI software, including the Qualcomm AI Engine Direct SDK and support for common runtimes.
The same block now appears in Qualcomm’s laptop chips and in its Dragonwing processors for industrial and embedded systems, which is where it matters for edge AI.
Inside the Hexagon NPU
Hexagon combines three kinds of engine that share local memory, so data can move between them without a trip to system RAM.
From DSP to NPU
Hexagon’s history explains why it is so efficient. Qualcomm has used Hexagon cores since the mid-2000s, first for modem and audio processing, where a phone must do continuous work on a tiny power budget. Around 2015, the Snapdragon 820 generation added Hexagon Vector eXtensions for camera and computer vision work. Dedicated tensor hardware followed in 2018 with the Snapdragon 855, aimed squarely at neural networks, and recent generations have enlarged the tensor engine and the shared memory around it.
Each step kept the original design goal: do the maths with as little data movement and as few watts as possible. That is the same requirement an edge AI system has when it runs on a vehicle supply or inside a sealed enclosure, which is why the design transfers well from phones to industrial boards.
-
//
Scalar cores. General control and the DSP heritage: first built for signal processing such as audio and camera data.
-
//
Vector engine. Hexagon Vector eXtensions (HVX), wide vector hardware for image processing and parts of a neural network that are not matrix maths.
-
//
Tensor engine. Dedicated matrix hardware for neural network layers, running INT8 and INT16, with INT4 and FP16 on newer generations.
-
//
Low power. Suits fanless, battery and vehicle systems, because the tensor work runs far more efficiently than on CPU cores.
Hexagon, Adreno and the CPU compared
Qualcomm describes the CPU, Adreno GPU and Hexagon NPU together as the Qualcomm AI Engine. Each can run part of an AI workload, and the software decides which one gets each model or layer.
| Engine | Main job | Best AI workloads | Typical precision | Power efficiency for AI |
|---|---|---|---|---|
| Kryo or Oryon CPU | Operating system and applications | Small models, pre- and post-processing | FP32, INT8 | Lowest |
| Adreno GPU | Graphics and display | Models with unusual layers, FP16 work | FP16, FP32 | Medium |
| Hexagon NPU | AI inference and signal processing | Quantized vision, audio and language models | INT8, INT16, INT4 | Highest |
In practice, a vision model quantized to INT8 runs on Hexagon, the CPU handles the camera pipeline and the application, and the GPU drives the display. Our NPU guide explains why a dedicated AI engine uses so much less power than CPU cores for the same model.
How software reaches Hexagon
Developers rarely program Hexagon directly. A trained model is converted and quantized, then run through one of Qualcomm’s runtimes, which places it on the NPU.
-
//
Qualcomm AI Engine Direct. Qualcomm’s main SDK for running models on Hexagon, the CPU and the GPU from one interface.
-
//
Qualcomm AI Hub. An online service with pre-optimised models and tools to compile and profile your own for a target chip.
-
//
Common runtimes. ONNX Runtime, LiteRT (formerly TensorFlow Lite) and ExecuTorch can hand work to Hexagon through Qualcomm back-ends.
-
//
Hexagon SDK. The low-level toolkit for writing custom DSP and vector code, used for signal processing rather than everyday AI.
Hexagon in industrial and embedded systems
Qualcomm processors bring phone-class power efficiency to edge AI. Dragonwing is Qualcomm’s name for its processors for industrial, robotics and IoT products, and each carries a Hexagon NPU. Performance scales across the range: the QCS6490 is rated at around 12 TOPS, while the IQ-9075 reaches up to 100 TOPS for multi-camera vehicle systems. Most run Linux or Android on Arm cores, and our guide to x86 vs Arm for industrial computers covers when that choice suits a project.
Our Qualcomm Dragonwing range covers the boards and systems we supply, from Qualcomm SBCs to in-vehicle AI computers with 16 GMSL camera inputs, including those built on the Qualcomm QCS6490.
Quick answers on Qualcomm Hexagon
Qualcomm Hexagon: common questions
Is Hexagon a DSP or an NPU?
Both. It began as a DSP and still does signal processing, but its added vector and tensor engines make it Qualcomm’s NPU. Qualcomm now calls it the Hexagon NPU.
How many TOPS does the Hexagon NPU deliver?
It depends on the chip. Qualcomm quotes figures per processor, not for Hexagon as a whole. Dragonwing processors in our range run from around 12 TOPS on the QCS6490 to up to 100 TOPS on the IQ-9075.
Can Hexagon run large language models?
Small ones, yes. Recent Hexagon generations run quantized language models with a few billion parameters on the device. Larger models need more memory than an embedded processor carries.
Does Hexagon work on Linux?
Yes. Dragonwing processors run Linux, including Qualcomm’s Yocto-based releases and Ubuntu builds, with the AI runtimes available on Linux as well as Android and Windows.
What is the Qualcomm AI Engine?
Qualcomm’s name for the CPU, GPU and Hexagon NPU together. It is a marketing term for the combined AI hardware, not a separate chip.
We build Qualcomm boards and systems with the operating system, AI runtime and storage fitted and configured as standard, and validate them before they ship. Tell us your models, cameras and power budget, and we will recommend the processor to match.