Double vs float datatypes

[pnbabu]

NEST-GPU uses a float datatype for the state and parameter arrays in the models. This was used instead of double to conserve memory in GPU and because often commercial GPUs do not support double datatype. However, simulations show that most of the memory is used for connections, not for the state and parameters.

In NEST-GPU, we could consider double datatypes for neuron state and parameters, and retain weight in connections as float. This requires more discussion with the NEST team and also a comparison with other neuromorphic hardware.