The same yolov4 engine on different Jetson cards

Hello,

I am working on two Jetson cards: AGX xavier dev kit and Xavier NX module.
and currently working with yolov4 on tensorrt, on both cards i have installed the latest jetpack version (4.5.1).

i have the following problem:
when i run my program on Xavier AGX dev kit, i get 33 FPS, while when i run it on the Xavier NX with the same engine from AGX dev kit i get 20FPS, when i try to generate an engine on the NX because it’s the best option to avoid problems and errors, my performance drops to less then 10 FPS. does anyone have an idea on why this is happening? is it safe to keep the same engine from the AGX since it runs faster.

Thank you.

Hi,

Could you try to enlarge the workspace size to see if it helps first?

$ /usr/src/tensorrt/bin/trtexec --onnx=[model] --workspace=2048

Thanks.

Hi,

I am not generating my engine with the onnx model but directly from the code without passing by the onnx model, the project it self is running directly on tensorrt and cuda, so unfortunatly i can’t change the workspace size.

Hi,

This is TensorRT API.
You can still set the workspace when creating an engine from API directly.

For example:
/usr/src/tensorrt/samples/sampleMNISTAPI/sampleMNISTAPI.cpp

// Build engine
builder->setMaxBatchSize(mParams.batchSize);
config->setMaxWorkspaceSize(16_MiB);
...

Thanks.