TX1 memory arrangement

I am trying to move mxnet to TX1.and run train_mnist successfully. Then I try to run some big cnn net(not so big).then an exception occurred:

terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc
  TX1 status:
KiB Mem :  4090604 total,  2847560 free,   583816 used,   659228 buff/cache
KiB Swap:        0 total,        0 free,        0 used.  3431420 avail Mem

There are still 2.8G free memory .It seems that GPU memory shortage causes the problem. How much memory can be used by Gpu ?

Ps.The Net’s parameter size is about 101M which I have trained and tested on pc.

 The problem is caused by application code and solved.