Memory usage increases substantially after unrolling the sequence, when starting the acquisition.

[LUMC-LowFieldMRI]

It looks like a copy of the unrolled sequence is created when the acquisition is started leading to excessive memory consumption.

[schote]

The sequence provider calculates the sequence block-wise and returns a list of numpy arrays.
In the current implementation, this list is concatenated before the sequence execution to obtain a single but very large numpy array. If I remember correctly, this step causes the increase in memory.
However, as part of the optimization of the sequence execution, we plan to start the sequence execution already during the sequence calculation (see #52). This requires to transfer the sequence data block-wise to the spectrum-card what potentially eliminates the problem of increasing memory.