I am responsible for a large project containing 577 Haskell library modules, with almost 120 KLOC. These modules are divided into packages. They can be compiled either (1) the old way, using ghc -M to generate dependency files, and then using make to call GHC once for each file. (2) the new way, compiling each package in one GHC invocation using --make. There is a big difference in speed between the two. The new way takes 4 minutes 16 seconds; the old way 6 minutes 55 seconds, both measured in wallclock time. (This is, admittedly, with -Onot.) My guess is that a lot of this is due to initialisation, particularly the need to read and parse lots of .hi files. This takes a lot of time, even though in fact all the .hi files during this test were on local disks. I think therefore the section "Sooner: producing a program more quickly" http://www.haskell.org/ghc/docs/latest/html/users_guide/sooner-faster-quicke... should recommend using --make to compile lots of files at a time. And by the way, that might not be a bad idea for GHC itself ...