TensorRT SSD high cpu memory usage

I used the sampleSSD code in tensorrt to compare and test the x86 tensorrt version and nx tensorrt7.1.3.0 version. Why does the same code have such a big difference in memory usage (the generated engine code is also the same), the version 5.0 occupies 1.0g ,the version 7.1 occupies 1.9g.
what are the possible optimization methods for nx ssd cpu memory usage?

Memory size problem need detailed analysis. But at high-level, it is possible that we may have changed some of kernels we used to improve performance and lead to memory increase. Also the lib size may also affect the usage.


Is there some possible optimization methods for nx ssd(tensorrt7.1.3) cpu memory usage?