ResNet-3D: Optimized for Qualcomm Devices

ResNet 3D is a network with 3D convolutions used for video understanding.

This is based on the implementation of ResNet-3D 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 1.24.3	Download
ONNX	w8a8	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
QNN_DLC	float	Universal	QAIRT 2.43	Download
QNN_DLC	w8a8	Universal	QAIRT 2.43	Download
TFLITE	float	Universal	QAIRT 2.43, TFLite 2.19.1	Download
TFLITE	w8a8	Universal	QAIRT 2.43, TFLite 2.19.1	Download

For more device-specific assets and performance metrics, visit ResNet-3D 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 ResNet-3D on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.video_classification

Model Stats:

• Model checkpoint: Kinetics-400
• Input resolution: 112x112
• Number of parameters: 33.4M
• Model size (float): 127 MB
• Model size (w8a8): 32.1 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
ResNet-3D	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	6.04 ms	2 - 201 MB	NPU
ResNet-3D	ONNX	float	Snapdragon® X2 Elite	7.065 ms	64 - 64 MB	NPU
ResNet-3D	ONNX	float	Snapdragon® X Elite	13.524 ms	63 - 63 MB	NPU
ResNet-3D	ONNX	float	Snapdragon® 8 Gen 3 Mobile	9.632 ms	0 - 255 MB	NPU
ResNet-3D	ONNX	float	Qualcomm® QCS8550 (Proxy)	13.194 ms	0 - 80 MB	NPU
ResNet-3D	ONNX	float	Qualcomm® QCS9075	25.967 ms	2 - 7 MB	NPU
ResNet-3D	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	7.903 ms	1 - 183 MB	NPU
ResNet-3D	ONNX	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	1.969 ms	0 - 174 MB	NPU
ResNet-3D	ONNX	w8a8	Snapdragon® X2 Elite	2.219 ms	33 - 33 MB	NPU
ResNet-3D	ONNX	w8a8	Snapdragon® X Elite	4.786 ms	32 - 32 MB	NPU
ResNet-3D	ONNX	w8a8	Snapdragon® 8 Gen 3 Mobile	3.335 ms	0 - 247 MB	NPU
ResNet-3D	ONNX	w8a8	Qualcomm® QCS6490	267.778 ms	56 - 137 MB	CPU
ResNet-3D	ONNX	w8a8	Qualcomm® QCS8550 (Proxy)	4.503 ms	1 - 3 MB	NPU
ResNet-3D	ONNX	w8a8	Qualcomm® QCS9075	4.737 ms	0 - 3 MB	NPU
ResNet-3D	ONNX	w8a8	Qualcomm® QCM6690	328.54 ms	43 - 49 MB	CPU
ResNet-3D	ONNX	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	2.76 ms	0 - 183 MB	NPU
ResNet-3D	ONNX	w8a8	Snapdragon® 7 Gen 4 Mobile	277.251 ms	59 - 67 MB	CPU
ResNet-3D	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	5.656 ms	2 - 209 MB	NPU
ResNet-3D	QNN_DLC	float	Snapdragon® X2 Elite	7.137 ms	2 - 2 MB	NPU
ResNet-3D	QNN_DLC	float	Snapdragon® X Elite	14.11 ms	2 - 2 MB	NPU
ResNet-3D	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	9.784 ms	0 - 284 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	91.631 ms	0 - 212 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	13.323 ms	2 - 4 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® SA8775P	24.044 ms	1 - 198 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® QCS9075	26.862 ms	2 - 6 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	28.113 ms	0 - 244 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® SA7255P	91.631 ms	0 - 212 MB	NPU
ResNet-3D	QNN_DLC	float	Qualcomm® SA8295P	26.388 ms	0 - 194 MB	NPU
ResNet-3D	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	7.826 ms	2 - 218 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	1.891 ms	1 - 64 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Snapdragon® X2 Elite	2.47 ms	1 - 1 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Snapdragon® X Elite	4.858 ms	1 - 1 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Snapdragon® 8 Gen 3 Mobile	3.388 ms	0 - 122 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® QCS6490	19.649 ms	1 - 3 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® QCS8275 (Proxy)	14.868 ms	1 - 60 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® QCS8550 (Proxy)	4.538 ms	1 - 136 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® SA8775P	4.631 ms	1 - 60 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® QCS9075	4.836 ms	1 - 3 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® QCM6690	91.01 ms	1 - 184 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® QCS8450 (Proxy)	6.227 ms	0 - 124 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® SA7255P	14.868 ms	1 - 60 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Qualcomm® SA8295P	8.284 ms	0 - 56 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	2.671 ms	0 - 178 MB	NPU
ResNet-3D	QNN_DLC	w8a8	Snapdragon® 7 Gen 4 Mobile	7.782 ms	1 - 169 MB	NPU
ResNet-3D	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	184.963 ms	0 - 218 MB	NPU
ResNet-3D	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	210.506 ms	0 - 277 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® QCS8275 (Proxy)	555.933 ms	0 - 219 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® QCS8550 (Proxy)	304.52 ms	0 - 2 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® SA8775P	277.219 ms	0 - 219 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® QCS9075	299.951 ms	0 - 69 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® QCS8450 (Proxy)	304.922 ms	1 - 276 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® SA7255P	555.933 ms	0 - 219 MB	NPU
ResNet-3D	TFLITE	float	Qualcomm® SA8295P	335.242 ms	0 - 213 MB	NPU
ResNet-3D	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	202.33 ms	0 - 219 MB	NPU
ResNet-3D	TFLITE	w8a8	Snapdragon® 8 Elite Gen 5 Mobile	166.104 ms	0 - 303 MB	NPU
ResNet-3D	TFLITE	w8a8	Snapdragon® 8 Gen 3 Mobile	192.525 ms	0 - 323 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® QCS6490	1743.476 ms	695 - 1060 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® QCS8275 (Proxy)	493.092 ms	0 - 252 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® QCS8550 (Proxy)	243.227 ms	0 - 2 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® SA8775P	246.162 ms	0 - 281 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® QCS9075	211.203 ms	0 - 67 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® QCM6690	1455.678 ms	660 - 817 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® QCS8450 (Proxy)	267.093 ms	0 - 321 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® SA7255P	493.092 ms	0 - 252 MB	NPU
ResNet-3D	TFLITE	w8a8	Qualcomm® SA8295P	256.386 ms	0 - 253 MB	NPU
ResNet-3D	TFLITE	w8a8	Snapdragon® 8 Elite For Galaxy Mobile	135.925 ms	0 - 252 MB	NPU
ResNet-3D	TFLITE	w8a8	Snapdragon® 7 Gen 4 Mobile	1164.19 ms	661 - 722 MB	NPU

License

• The license for the original implementation of ResNet-3D can be found here.

References

• A Closer Look at Spatiotemporal Convolutions for Action Recognition
• 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.