The ailia SDK, an inference framework for edge devices, uses the ONNX format to perform fast inference on the GPU. In this article, I introduce some findings regarding ONNX model optimization obtained in the process of developing the ailia SDK.