The aim is to reduce HTTP requests. In addition, sometimes the compressed sprite will weigh less than the original images.
Recently I had a website with a lot of transparent gradients (white to trans, grey to trans) and some black and white on transparent images. By putting them all in a sprite and reducing the colors in the png to 8 I could get the sprite to be smaller in filesize than the original images (just… it was about a 0.5% saving). By reducing the number of HTTP requests from 10 to 1 though meant the site loaded faster (if measuring time from first connection to all data transferred).
In that case, a measurable increase was found.
I agree though that it’s possible to mess things up and to end up with a larger sprite than needed, especially if you aren’t using PNG compression.
Note two years after posting this – if you are using SSL, you should look into SPDY (my note in a further two years will mention HTTP 2.0 instead of SPDY!). SPDY negates the benefits of spriting.