I don't know much about Scheme (or Lisp at all really; I've only become interested in it during the last month or so) but it seems that the Scheme compilers produce quite sluggish code, at least looking through the grainy, distorted lens that is the Computer Language Benchmarks Game.
From some of the microbenchmarks published by the Gambit guy(s) on their homepage, it looks as if it fares really well against the Chicken and Mzscheme compilers, but I wanted to examine its performance on problems relative to SBCL, which is the Common Lisp compiler I've arbitrarily chosen.
Since it's not featured on the language shootout, I pulled the SBCL code for a few tests there as well as the Gambit code which was already prepared on their site (why hasn't Gambit appeared on the shootout site for the nosy hackers who want to find out if less mainstream (man I hate that word) languages are ridiculously slow? Many of them aren't! And they're more expressive and comfortable than C++/Java, use them!) and ran a few comparisons.
SBCL comes out very much on top; in general the Gambit programs take two or three times as long to do their job (although I haven't looked at memory usage). But as far as Scheme compilers go, Gambit seems to be improving things.
(Really dodgy) timings on my Macbook (input values = whatever's on the shootout site):
One thing I've noticed from these coarse measurements is the "sys" component of the execution time is mostly higher for SBCL; maybe this means a lot of allocation and de-allocation of stuff is going on ...Or something.