Tensor Cores and deep learning
10 views (last 30 days)
Gerardo Suarez on 24 Jun 2019
I have a GeForce RTX 2060. How can I take advantage of its 240 tensor cores for deep learning? I have read that I should set the precision to FP16, but using the "trainnetwork" function there does not seem to be any option.
Krishna Bindumadhavan on 14 Sep 2019
Edited: Krishna Bindumadhavan on 14 Sep 2019
Currently, there is no support for training a neural network with half precision data type. This is expected in a future release.
You can however take advantage of tensor cores, while generating code for deploying a trained neural network with GPU Coder. You can look at Deep Learning Prediction with NVIDIA TensorRT example , but set the DataType property of the DeepLearningConfig to 'fp16'. This will use the Tensor cores on a Volta or Turing card such as the RTX series for inference.
In addition, MATLAB supports the half data type via the half precision object in the fixed point designer toolbox: https://www.mathworks.com/help/fixedpoint/ref/half.html?s_tid=doc_ta. General code generation support for the half type, via MATLAB Coder and GPU coder is under active development and can be expected in an upcoming release.