Ideally you want a non-redundant sequence set for this task, that's why NR is the default choice for many applications. A smaller alternative would be UniRef50, which still takes about 25 GB (see https://www.biostars.org/p/9591499/#9591570). If computer resources are the problem you could even run PSI-BLAST remotely (see https://www.biostars.org/p/44096/), but of course that takes some time for each query.

SwissProt is sparse as it curates some organisms much deeper than others. For instance, among plants, it contains ~16K reviewed proteins from model Arabidopsis thaliana, while it contains only 46 of Brachypodium distachyon, another model. This will taxonomically bias sequence searches of SwissProt. However, according to https://www.uniprot.org/help/redundancy, SwissProt contains "one record per gene in one species", so it is non-redundant in that sense. That's probably why some tools can use it to compute PSSMs, see for instance https://toolkit.tuebingen.mpg.de/tools/hmmer.