# Qualcomm® AI Engine Direct Backend Library
The TFLite Delegate is built on the top of Qualcomm® AI Engine Direct libraries, which can be found in
the corresponding architecture directory in the Qualcomm® AI Engine Direct SDK `lib` directory.
To support the widest range of devices, we recommend to put all libraries to the
environment. However, if the application size is a concern, here are descriptions
for each library:
| Backend Name | Backend Description | Target and Library Names | Library Description |
| --- | --- | --- | --- |
| GPU | Backend for Adreno™ GPU hardware accelerator |
| **libQnnGpu.so**
GPU backend library, same name used across all devices. |
| DSP | Backend for Hexagon™ DSP hardware accelerator in SM8150 or SM8250. |
hexagon-v66
aarch64-android
libQnnDsp.so
libQnnDspV66Stub.so
| **libQnnDsp.so**
This library loads corresponding Stub libraries according to the runtime environment.
Then the Stub library on CPU side loads corresponding Skel library on DSP side.
V66 libraries are loaded in S865.
**libQnnDspV66Skel.so**
DSP native library loaded by `libQnnDspV66Stub.so`
**libQnnDspV66Stub.so**
DSP proxy backend library on CPU side, communicating with `libQnnDspV66Skel.so` |
| HTP | Backend for Hexagon™ HTP hardware accelerator in S888 or SoC newer than S888. |
hexagon-v68
hexagon-v69
hexagon-v73
hexagon-v75
hexagon-v79
aarch64-android
libQnnHtp.so
libQnnHtpPrepare.so
libQnnHtpV68Stub.so
libQnnHtpV69Stub.so
libQnnHtpV73Stub.so
libQnnHtpV75Stub.so
libQnnHtpV79Stub.so
| **libQnnHtp.so**
There are two variants of this library: (Android) This library loads corresponding Stub libraries according to the runtime environment.
Then the Stub library on CPU side loads corresponding Skel library on HTP side. (x86\_64) Serves as an emulator for the hardware accelerator.
**libQnnHtpV\*Skel.so**
HTP native library loaded by corresponded `libQnnHtpV*Stub.so`
**libQnnHtpV\*Stub.so**
HTP porxy library on CPU side, communicating with `libQnnHtpV*Skel.so`
**libQnnHtpPrepare.so**
HTP library run on CPU side. This library is loaded to prepare the graph.
However, if the cache exists and valid for a certain model, this library
might not need to be on the device. |
| System | Library used by Qualcomm® AI Engine Direct to interpret the cache. | | **libQnnSystem.so**
If the cache feature is used, this library is required on the device to
save or interpret the cache file correctly. |
Last Published: Oct 10, 2025
[Previous Topic
External Delegate Options](https://docs.qualcomm.com/bundle/publicresource/80-63442-50/topics/options.md) [Next Topic
Release Notes](https://docs.qualcomm.com/bundle/publicresource/80-63442-50/topics/release_notes.md)