Currently the GPU implementation of training does not parallelize for single series. If you have a long series and can break it up into many shorter series, then it will parallelize well.
For instance, if your series looks like this, it will run on only one core which will be slow:
input dimensions = num_inputs-by-timesteps cell of Nx1 values
target dimensions = num_outputs-by-timesteps cell of Mx1 values
If you have multiple series, your data will look like this, where Q is the number of series being trained on or evaluated in parallel:
input dimensions = num_inputs-by-timesteps cell of NxQ values
target dimensions = num_outputs-by-timesteps cell of MxQ values
The larger the number of series Q, the more parallelism on the GPU.
