Deep learning with a GPU that supports fp16

Hi.

NVDIA has released the new RTX 2XXX and 3XXX series that support fp16 that accelrates training process.

Does Matlab support this?

Thank you

[Walter Roberson]

According to the release notes it does; https://www.mathworks.com/products/gpu-coder/whatsnew.html

but according to the Product Limitations it does not.

https://www.mathworks.com/help/gpucoder/ug/gpucoder-limitations.html

[Joss Knight]

It is supported for deep learning code generation, but not for general code generation.

[Walter Roberson]

An interesting article came through recently, https://www.linkedin.com/pulse/deep-learning-cant-progress-ieee-754-floating-point-heres-omtzigt/

[Krishna Bindumadhavan]

There is support for half precision in MATLAB via the half precision object, available in the fixed point designer toolbox:https://www.mathworks.com/help/fixedpoint/ref/half.html.

General Code generation support for half precision data type via MATLAB Coder and GPU Coder is under active development. This functionality is expected in an upcoming release.

As mentioned below, there is no support currently for using half for training a deep learning network in MATLAB. This is expected in a future release.

Answer 1

[Joss Knight]

You can take advantage of FP16 when generating code for prediction on a deep neural network. Follow the pattern of 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 a Volta or Turing card such as the RTX series.

There is no way yet to use half precision or Tensor cores for training a deep neural network in MATLAB. This is expected in an upcoming release.

[J. Womack]

Is this released yet?

[Joss Knight]

Training in half precision is not released yet.

[Juuso Korhonen]

What about now? Or do we have to wait for 2021 release?

[Joss Knight]

You can use the Deep Network Quantizer to calibrate a trained network for 8-bit reduced precision types. For now, fp16 is not supported, and quantization-aware training is not supported.

With an Ampere card, using the latest R2021a release of MATLAB (soon to be released), you will be able to take advantage of the Tensor cores using single precision because of the new TF32 datatype that cuDNN leverages when performing convolutions on an Ampere card.