ResNeXt50: Optimized for Qualcomm Devices

ResNeXt50 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 ResNeXt50 found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up 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 2.42, ONNX Runtime 1.24.1	Download
ONNX	w8a8	Universal	QAIRT 2.42, ONNX Runtime 1.24.1	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a8	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.17.0	Download
TFLITE	w8a8	Universal	QAIRT 2.43, TFLite 2.17.0	Download

For more device-specific assets and performance metrics, visit ResNeXt50 on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models 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 ResNeXt50 on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.image_classification

Model Stats:

• Model checkpoint: Imagenet
• Input resolution: 224x224
• Number of parameters: 25.0M
• Model size (float): 95.4 MB
• Model size (w8a8): 24.8 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
ResNeXt50	ONNX	float	Snapdragon® X Elite	2.417 ms	50 - 50 MB	NPU
ResNeXt50	ONNX	float	Snapdragon® 8 Gen 3 Mobile	1.694 ms	0 - 146 MB	NPU
ResNeXt50	ONNX	float	Qualcomm® QCS8550 (Proxy)	2.279 ms	1 - 10 MB	NPU
ResNeXt50	ONNX	float	Qualcomm® QCS9075	3.487 ms	1 - 4 MB	NPU
ResNeXt50	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	1.384 ms	0 - 84 MB	NPU
ResNeXt50	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	1.143 ms	1 - 86 MB	NPU
ResNeXt50	ONNX	float	Snapdragon® X2 Elite	1.09 ms	50 - 50 MB	NPU
ResNeXt50	ONNX	w8a8	Snapdragon® X Elite	1.264 ms	25 - 25 MB	NPU
ResNeXt50	ONNX	w8a8	Snapdragon® 8 Gen 3 Mobile	0.822 ms	0 - 100 MB	NPU
ResNeXt50	ONNX	w8a8	Qualcomm® QCS6490	46.93 ms	7 - 24 MB	CPU
ResNeXt50	ONNX	w8a8	Qualcomm® QCS8550 (Proxy)	1.093 ms	0 - 30 MB	NPU
ResNeXt50	ONNX	w8a8	Qualcomm® QCS9075	1.23 ms	0 - 3 MB	NPU
ResNeXt50	ONNX	w8a8	Qualcomm® QCM6690	26.196 ms	3 - 13 MB	CPU
ResNeXt50	ONNX	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	0.699 ms	0 - 78 MB	NPU
ResNeXt50	ONNX	w8a8	Snapdragon® 7 Gen 4 Mobile	21.717 ms	4 - 13 MB	CPU
ResNeXt50	ONNX	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	0.598 ms	0 - 79 MB	NPU
ResNeXt50	ONNX	w8a8	Snapdragon® X2 Elite	0.518 ms	25 - 25 MB	NPU
ResNeXt50	QNN_DLC	float	Snapdragon® X Elite	2.673 ms	1 - 1 MB	NPU
ResNeXt50	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	1.804 ms	0 - 142 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	11.9 ms	1 - 75 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	2.541 ms	1 - 2 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® SA8775P	3.807 ms	1 - 76 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® QCS9075	3.659 ms	3 - 5 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	4.05 ms	0 - 115 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® SA7255P	11.9 ms	1 - 75 MB	NPU
ResNeXt50	QNN_DLC	float	Qualcomm® SA8295P	4.116 ms	0 - 53 MB	NPU
ResNeXt50	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	1.452 ms	0 - 76 MB	NPU
ResNeXt50	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	1.186 ms	1 - 77 MB	NPU
ResNeXt50	QNN_DLC	float	Snapdragon® X2 Elite	1.458 ms	1 - 1 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Snapdragon® X Elite	1.259 ms	0 - 0 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Snapdragon® 8 Gen 3 Mobile	0.815 ms	0 - 95 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® QCS6490	3.089 ms	2 - 4 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® QCS8275 (Proxy)	2.481 ms	0 - 68 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® QCS8550 (Proxy)	1.094 ms	0 - 2 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® SA8775P	1.486 ms	0 - 70 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® QCS9075	1.233 ms	0 - 2 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® QCM6690	7.631 ms	0 - 193 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® QCS8450 (Proxy)	1.46 ms	0 - 98 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® SA7255P	2.481 ms	0 - 68 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Qualcomm® SA8295P	1.791 ms	0 - 66 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	0.625 ms	0 - 67 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Snapdragon® 7 Gen 4 Mobile	1.393 ms	0 - 73 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	0.537 ms	0 - 73 MB	NPU
ResNeXt50	QNN_DLC	w8a8	Snapdragon® X2 Elite	0.703 ms	0 - 0 MB	NPU
ResNeXt50	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	1.764 ms	0 - 186 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® QCS8275 (Proxy)	11.939 ms	0 - 122 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® QCS8550 (Proxy)	2.484 ms	0 - 2 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® SA8775P	16.126 ms	0 - 123 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® QCS9075	3.703 ms	0 - 52 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® QCS8450 (Proxy)	3.979 ms	0 - 161 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® SA7255P	11.939 ms	0 - 122 MB	NPU
ResNeXt50	TFLITE	float	Qualcomm® SA8295P	4.158 ms	0 - 104 MB	NPU
ResNeXt50	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	1.427 ms	0 - 128 MB	NPU
ResNeXt50	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	1.19 ms	0 - 126 MB	NPU
ResNeXt50	TFLITE	w8a8	Snapdragon® 8 Gen 3 Mobile	0.671 ms	0 - 96 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® QCS6490	2.868 ms	0 - 27 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® QCS8275 (Proxy)	2.224 ms	0 - 67 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® QCS8550 (Proxy)	0.92 ms	0 - 2 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® SA8775P	1.289 ms	0 - 68 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® QCS9075	1.013 ms	0 - 27 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® QCM6690	7.348 ms	0 - 190 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® QCS8450 (Proxy)	1.264 ms	0 - 97 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® SA7255P	2.224 ms	0 - 67 MB	NPU
ResNeXt50	TFLITE	w8a8	Qualcomm® SA8295P	1.564 ms	0 - 62 MB	NPU
ResNeXt50	TFLITE	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	0.561 ms	0 - 71 MB	NPU
ResNeXt50	TFLITE	w8a8	Snapdragon® 7 Gen 4 Mobile	1.236 ms	0 - 67 MB	NPU
ResNeXt50	TFLITE	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	0.495 ms	0 - 70 MB	NPU

License

• The license for the original implementation of ResNeXt50 can be found here.

References

• Aggregated Residual Transformations for Deep Neural Networks
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.