In the landscape you sketch, I bet profiling will show the overhead being at the DB (DBI + DBD + database interface) side of the program, and not in Text::CSV_XS. Yes, Text::CSV_XS does have the SV overhead, but when you use bind_columns, most of that overhead is neglectable (comparable to other overhead).

You didn't state what database you use, and which interface and/or connection method. Sometimes specific options and attributes can really speed up your process.

We really need more in order to be able to help you (if possible at all).