DLA makes inference much slower

Hello, I tried to enable DLA on my inference program, but it runs much slower than without DLA. Can anyone help?

I am using Jetson Xavier, and tested DLA on my own program, also sampleMNIST with the following commands:

./sample_mnist --fp16 --useDLACore=1
./sample_mnist --fp16 --useDLACore=0
./sample_mnist --fp16