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.
NEST-GPU uses a
floatdatatype for the state and parameter arrays in the models. This was used instead ofdoubleto conserve memory in GPU and because often commercial GPUs do not supportdoubledatatype. However, simulations show that most of the memory is used for connections, not for the state and parameters.In NEST-GPU, we could consider
doubledatatypes for neuron state and parameters, and retainweightin connections asfloat. This requires more discussion with the NEST team and also a comparison with other neuromorphic hardware.