[TensorRT] Why is my converted engine faster than the saved engine?

Hello,

i use my cnn(Trained with caffe) with TensorRT and i load it via:

engine = trt.utils.caffe_to_trt_engine(G_LOGGER,
                                           MODEL_PROTOTXT,
                                           CAFFE_MODEL,
                                           1,
                                           256 << 20,
                                           OUTPUT_LAYERS,
                                           trt.infer.DataType.FLOAT)

or

engine2 = trt.utils.load_engine(G_LOGGER,'/path/to/cnn.engine')

the second one i saved with:

trt.utils.write_engine_to_file('/path/to/cnn.engine', engine.serialize())

the first option is about 20 to 30 percent faster and i don’t know why.
Does anybody have a idea why this happens?

Thanks!