No 46000 is not a lot of sequences and there is nothing wrong with aligning the whole lot. Furthermore, 6000bp is not large.

Your idea is correct, but you are inferring a file size of 4600 sequences at 6000bp. The limitation is either the multiprocessing or the computer, e.g. low RAM. You shouldn't need to shrink the data sets so much. How many cores do you have on your machine?

I used the Clustal family for sometime but switched for the sort of issues you are encountering. It is possible Clustal O is part an issue, there are better (below).

Anyway, simply google FAMSA. You need to keep in mind that both speed and accuracy are criteria.

My personal advice is align under MAFFT or FAMSA and switch the formats to Clustal. You can do this via coding (Biopython) or simply load and save in Clustal to get the correct format.

Clustal Omega was upgraded in 2018 and that publication is worth reading, it was fairly astonishing that not until that point the clustalo implement specific aligning for DNA, previously it was shoe-horned from proteins. They also euloguise the developments of Muscle for alignment accuracy which they implemented.

MAFFT is the aligner that is generally recognised for both speed and accuracy now, if you look at the FAMSA and Clustal Omega publication of their benchmarks you'll find that MAFFT performs very well.

FAMSA is built for speed, it is however for coded for proteins and DNA is shoe-horned. FAMSA is fast and multiprocesses very well. It claims to retain accuracy over large alignments, whereas its benchmarks show other alignment algorithms including MAFFT and Clustal O drop in performance accuracy.

I would say its down to two alignment programs. There is still room for improvement however, I code around this area at present.

If you are sticking with Clustal O, then its maximising the margins, thus multiprocessing is essential using a machine with maximum cores. Its certainly doable, I would look at increasing the number of sequences per file however, like 2 files at 6000bp. Minimising 'redundancy' depends on what your question is, because the frequency of a given allele might be part of the investigation

ClustalO should be able to handle 7525 x 6000 output, but I would ask myself whether I needed all those sequences. It is unlikely that you have 7525 completely unique L1s, or that you truly want to inspect 7525 L1s even if they were unique. Think about it: how would you visualize an alignment that is 7525 x 6000 characters? Assuming that you have short sequence headers and that there are not very many indels, and let's say 3000 characters per page, that would still be in excess of 15000 pages. Even if you were to convert this into a phylogenetic tree, the output would still be in hundreds of pages.

But don't worry, many people ask about aligning thousands or tens of thousands of sequences that are even longer than yours, without thinking how exactly they are going to look at it. My point is that a smaller number of carefully chosen sequences is usually easier to inspect, and carries the same information. That means clustering at various identity thresholds, such as 99%, 95% or even 90%. I suspect that the latter would cut down your sequence number at least in half, which is probably still too many. CD-HIT will quickly cluster both protein and DNA sequences and help remove the redundancy.

Starting from a redundant dataset, I suggest you run clustalo with all the threads you have, and make sure to specify --seqtype=DNA --infmt=fa --outfmt=clu --threads=8 (use actual number) --verbose in addition to input and output switches. The verbose switch will give you an update as it goes so hopefully you have some idea when it gets close to the end. It will help if you have lots of RAM, many threads and a speedy computer.