---
library_name: pytorch
license: other
tags:
- backbone
- bu_auto
- android
pipeline_tag: image-classification

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/web-assets/model_demo.png)

# GPUNet: Optimized for Qualcomm Devices

GPUNet is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.

This is based on the implementation of GPUNet found [here](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/GPUNet).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/gpunet) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device.

## Getting Started
There are two ways to deploy this model on your device:

### Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-onnx-float.zip)
| ONNX | w8a16 | Universal | QAIRT 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-onnx-w8a16.zip)
| ONNX | w8a8 | Universal | QAIRT 2.37, ONNX Runtime 1.23.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-qnn_dlc-float.zip)
| QNN_DLC | w8a16 | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-qnn_dlc-w8a16.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.42 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.42, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.42, TFLite 2.17.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/gpunet/releases/v0.46.0/gpunet-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[GPUNet on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/gpunet)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/gpunet) Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for [GPUNet on GitHub](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/gpunet) for usage instructions.

## Model Details

**Model Type:** Model_use_case.image_classification

**Model Stats:**
- Model checkpoint: Imagenet
- Input resolution: 224x224
- Number of parameters: 10.49M
- Model size (float): 45.28MB
- Model size (w8a8): 21.3MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| GPUNet | ONNX | float | Snapdragon® X Elite | 1.12 ms | 24 - 24 MB | NPU
| GPUNet | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 0.955 ms | 0 - 126 MB | NPU
| GPUNet | ONNX | float | Qualcomm® QCS8550 (Proxy) | 1.231 ms | 0 - 198 MB | NPU
| GPUNet | ONNX | float | Qualcomm® QCS9075 | 1.509 ms | 1 - 3 MB | NPU
| GPUNet | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.739 ms | 0 - 99 MB | NPU
| GPUNet | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.636 ms | 0 - 99 MB | NPU
| GPUNet | ONNX | w8a16 | Snapdragon® X Elite | 0.974 ms | 12 - 12 MB | NPU
| GPUNet | ONNX | w8a16 | Snapdragon® 8 Gen 3 Mobile | 0.777 ms | 0 - 131 MB | NPU
| GPUNet | ONNX | w8a16 | Qualcomm® QCS6490 | 102.34 ms | 18 - 33 MB | CPU
| GPUNet | ONNX | w8a16 | Qualcomm® QCS8550 (Proxy) | 0.995 ms | 0 - 25 MB | NPU
| GPUNet | ONNX | w8a16 | Qualcomm® QCS9075 | 1.223 ms | 0 - 3 MB | NPU
| GPUNet | ONNX | w8a16 | Qualcomm® QCM6690 | 53.444 ms | 28 - 36 MB | CPU
| GPUNet | ONNX | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 0.556 ms | 0 - 114 MB | NPU
| GPUNet | ONNX | w8a16 | Snapdragon® 7 Gen 4 Mobile | 40.754 ms | 30 - 37 MB | CPU
| GPUNet | ONNX | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 0.463 ms | 0 - 114 MB | NPU
| GPUNet | ONNX | w8a8 | Snapdragon® X Elite | 0.713 ms | 12 - 12 MB | NPU
| GPUNet | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.583 ms | 0 - 125 MB | NPU
| GPUNet | ONNX | w8a8 | Qualcomm® QCS6490 | 16.722 ms | 4 - 18 MB | CPU
| GPUNet | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.827 ms | 0 - 26 MB | NPU
| GPUNet | ONNX | w8a8 | Qualcomm® QCS9075 | 0.947 ms | 0 - 3 MB | NPU
| GPUNet | ONNX | w8a8 | Qualcomm® QCM6690 | 10.135 ms | 11 - 19 MB | CPU
| GPUNet | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.507 ms | 0 - 107 MB | NPU
| GPUNet | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 7.572 ms | 9 - 17 MB | CPU
| GPUNet | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.469 ms | 0 - 113 MB | NPU
| GPUNet | QNN_DLC | float | Snapdragon® X Elite | 1.391 ms | 1 - 1 MB | NPU
| GPUNet | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 0.947 ms | 0 - 62 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 4.721 ms | 1 - 33 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 1.297 ms | 1 - 105 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® SA8775P | 1.713 ms | 1 - 36 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® QCS9075 | 1.569 ms | 1 - 3 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 2.408 ms | 0 - 64 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® SA7255P | 4.721 ms | 1 - 33 MB | NPU
| GPUNet | QNN_DLC | float | Qualcomm® SA8295P | 2.283 ms | 0 - 33 MB | NPU
| GPUNet | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.707 ms | 0 - 32 MB | NPU
| GPUNet | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.565 ms | 1 - 37 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Snapdragon® X Elite | 1.234 ms | 0 - 0 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Snapdragon® 8 Gen 3 Mobile | 0.768 ms | 0 - 57 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® QCS6490 | 3.257 ms | 2 - 4 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® QCS8275 (Proxy) | 2.494 ms | 0 - 41 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® QCS8550 (Proxy) | 1.065 ms | 0 - 2 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® SA8775P | 1.293 ms | 0 - 43 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® QCS9075 | 1.226 ms | 2 - 4 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® QCM6690 | 6.566 ms | 0 - 160 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® QCS8450 (Proxy) | 1.458 ms | 0 - 61 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® SA7255P | 2.494 ms | 0 - 41 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Qualcomm® SA8295P | 1.656 ms | 0 - 39 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Snapdragon® 8 Elite For Galaxy Mobile | 0.528 ms | 0 - 40 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Snapdragon® 7 Gen 4 Mobile | 1.284 ms | 0 - 42 MB | NPU
| GPUNet | QNN_DLC | w8a16 | Snapdragon® 8 Elite Gen 5 Mobile | 0.434 ms | 0 - 45 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Snapdragon® X Elite | 0.732 ms | 0 - 0 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.452 ms | 0 - 56 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 1.969 ms | 0 - 2 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 1.427 ms | 0 - 39 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.607 ms | 0 - 3 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® SA8775P | 0.797 ms | 0 - 40 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 0.691 ms | 0 - 2 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 3.448 ms | 0 - 42 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 0.846 ms | 0 - 58 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® SA7255P | 1.427 ms | 0 - 39 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Qualcomm® SA8295P | 1.073 ms | 0 - 37 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.338 ms | 0 - 37 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.794 ms | 0 - 38 MB | NPU
| GPUNet | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.294 ms | 0 - 44 MB | NPU
| GPUNet | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 0.941 ms | 0 - 97 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 4.701 ms | 0 - 63 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 1.275 ms | 0 - 3 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® SA8775P | 7.192 ms | 0 - 63 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® QCS9075 | 1.574 ms | 0 - 27 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 2.401 ms | 0 - 93 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® SA7255P | 4.701 ms | 0 - 63 MB | NPU
| GPUNet | TFLITE | float | Qualcomm® SA8295P | 2.255 ms | 0 - 62 MB | NPU
| GPUNet | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 0.713 ms | 0 - 61 MB | NPU
| GPUNet | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 0.564 ms | 0 - 66 MB | NPU
| GPUNet | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.34 ms | 0 - 54 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® QCS6490 | 1.649 ms | 0 - 15 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 1.12 ms | 0 - 37 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 0.422 ms | 0 - 2 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® SA8775P | 0.62 ms | 0 - 40 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® QCS9075 | 0.524 ms | 0 - 14 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® QCM6690 | 2.998 ms | 0 - 39 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 0.685 ms | 0 - 55 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® SA7255P | 1.12 ms | 0 - 37 MB | NPU
| GPUNet | TFLITE | w8a8 | Qualcomm® SA8295P | 0.865 ms | 0 - 35 MB | NPU
| GPUNet | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.276 ms | 0 - 36 MB | NPU
| GPUNet | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 0.627 ms | 0 - 36 MB | NPU
| GPUNet | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 0.237 ms | 0 - 43 MB | NPU

## License
* The license for the original implementation of GPUNet can be found
  [here](http://www.apache.org/licenses/LICENSE-2.0).

## References
* [GPUNet: Searching the Deployable Convolution Neural Networks for GPUs](https://arxiv.org/abs/2205.00841)
* [Source Model Implementation](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/GPUNet)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).