You should reduce redundancy since it is unlikely that all 11K genomes are completely unique w.r.t sequence. SARS genomes?

A multiple sequence alignment of that many sequences which are that long is highly likely to be spurious, unless all the genomes are incredibly similar (in which case you might as well remove redundant identical sequences).

On the lines of what genomax and Joe already alluded to, you can reduce redundancy in 2 ways:

1. At the whole genome sequence level (tracking which genomes are identical at even 100% sequence identity), with CD-HIT or some such tool,
2. but regardless, for the MAFFT alignment, you can remove uninformative alignment columns with a tool like here - https://docs.qiime2.org/2020.2/plugins/available/phylogeny/align-to-tree-mafft-fasttree/ OR one of the tools listed in this comparison of methods paper (see Table 2 ) - https://academic.oup.com/mbe/article/30/3/689/1040880

And submit this reduced representation to RaXML at the CIPRES Gateway, that might help...