Aurora 1.5 Ensemble has the option of being based on an ensemble with N members. This is currently achieved by running the model N times in the context in which the ensemble approach is required.
There should be an option to make the ensemble members internal to Aurora, so that getting all of the ensemble results is done using a single call to e.g. forward, as opposed to running N of them in a loop.
This would expose more opportunities for making the full use of GPUs. Currently, each ensemble member is a completely separate model with separate inputs and outputs. This means that a high-performance GPU may be under-utilised compared to representing all ensemble members using the same tensor.
Aurora 1.5 Ensemble has the option of being based on an ensemble with
Nmembers. This is currently achieved by running the modelNtimes in the context in which the ensemble approach is required.There should be an option to make the ensemble members internal to Aurora, so that getting all of the ensemble results is done using a single call to e.g.
forward, as opposed to runningNof them in a loop.This would expose more opportunities for making the full use of GPUs. Currently, each ensemble member is a completely separate model with separate inputs and outputs. This means that a high-performance GPU may be under-utilised compared to representing all ensemble members using the same tensor.