That warning does pop up every now and then. Basically, if you try to filter an excessively large number of tracks down to an excessively small number of tracks, eventually each track you remove provides an insignificant improvement to the model, and the discrete nature of a streamlines reconstruction (i.e. individual trajectories rather than a ‘field’ of connectivity) begins to have an effect on the algorithm’s performance - this is what I refer to in the SIFT paper as the ‘quantisation limit’ (precise details in the manuscript). All that warning message means is that you’ve passed that limit, but because you specifically asked the command to reach a certain number of streamlines before terminating, it is proceeding anyway.
The amount of RAM required for SIFT depends on the input number of tracks and the image resolution (and also a little on the streamlines seeding mechanism used). There’s a table in the SIFT2 paper showing execution times and RAM requirements for a few different use cases. At 2.5mm isotropic, the requirement is about 200-300MB per million streamlines.
For fixelcfestats, running on an upsampled (1.25mm) template requires the better half of a 128GB machine.