Along my basic tutorial develop, I recall the ZYNQ itself cannot handle that much of computation.
As the DDR3 itself only got 1GB. So what would you expect is quite expectable.
So I will highly suggest train on host and inference on ZYNQ.
Meantime, if you are lazy as me you can use the weight load function only via the exported weight file.
it is been a while, so any new ARM-based low tensorlite might make me wrong.
