TensorRT consumes almost the same GPU memory between int8 and float32

Question

I use TensorRT engine to speedup the DNN inference. I found TensorRT consumes almost the same GPU memory between int8 and float32. Does the feature map use float32 store mo