Antonio.Aubry , just to put this somewhat into perspective, I recently used both salmon and hisat2 (the latter being a common RNA_seq aligner) on a larger number of samples (157 RNA-seq samples, 2x25bp, total of 1.7 bio paired-end reads) using one 72-core Skylake node. For salmon using 8 threads and running 4 jobs in parallel it took about 1h12min. The same for hisat2, 12 threads and piped into samtools sort with 4 jobs in parallel took 3h30min. The latter might become a bit faster without sorting the BAM file, but I think you get the idea: Salmon is pretty fast so especially suited for desktop PCs where resources are rather limited.

I would typically run TopHat2 with 8 GB of RAM and 4 cores on a cluster (and I ran STAR with similar run time a few times, even though I think it tends to use more RAM when run on a computer that has extra RAM available). However, that is with 50 bp Single-End reads: if you have paired-end reads (and want to focus more on splicing and/or mutation calling), you may need additional resources and/or time (but that also means you'll need to do a genome alignment).

I guess this really relates to Brett's answer suggesting considering use of a compute cluster, but I thought I should also say something here.

Running a transcription quantification (like Salmon or kallisto) on a local computer shouldn't be a problem.

If you had a MiSeq/MiniSeq/iSeq experiments with a lower number of total reads in a SE 50 bp polyA library (say 5-10M reads per sample), you might even be able to do a genome alignment in serial (assuming you have a 2-group comparison with triplicates) on a computer with 8 GB of RAM and 2-4 cores within one day. In that case, I do think there are benefits to being able to visualize your alignment. However, most people would probably use a cluster for genome alignments (and probably have more reads from a NextSeq/HiSeq). So, learning to submit jobs on a cluster is also a very useful skill :)