You may take advantage of tensor cores after generating code for a trained neural network with GPU Coder. You can take a look at the Deep Learning Prediction with NVIDIA TensorRT example but set the DataType property of the DeepLearningConfig to 'fp16'. This will use the Tensor cores on the Titan-V.
In MATLAB, there is support for the half precision data type via the half precision object in the fixed point designer toolbox:
In addition, general code generation support for fp16 data type via MATLAB Coder and GPU coder is under active development and can be expected in a future release.
