I think Corion has the right idea, but I'd take it one step further and avoid having the perl script write directly to disk.

I'd do it this way:

gunzip -cd big.tsv.gz 
| perl -ne"@v=split chr(9),$_; $i=-1; print $v[$i+=3], chr(9) while $i
+ < @v; print chr(10)" 
| perl -pe1 > newfile
[download]

NB: That's a one-liner wrapped across 3 lines for posting.

The idea is to

1. Minimise the memory juggling inside the first perl script.

   By avoiding building a second array of the columns you are keeping and then more memory in order to do the join.

2. Use the pipe buffers between the first perl and the second to buffer the IO and prevent some disk contention.

If you are going to be doing this regularly rather than just as a one-off, it might be worth trying both ways to see which works best on your system.