I upgraded my cable modem at home to 6.0 Mb down/512 Kb up, and ran
some tests on a 4501 with 1.2b3 vs. 1.2b7 vs. pfSense.  Though my
methodology isn't perfect, this was a quick test and I can repeat the
same results over and over again.
http://chrisbuechler.com/4vs5/

In addition to the stats Manuel posted here:
http://m0n0.ch/wall/list/?action=show_msg&actionargs[]=151&actionargs[]=57

In short, 1.2b3 on a 4501 could easily handle my full 6 Mb down.  If
the meter on the front page is to be believed (it's not extremely
accurate, but I don't know exactly how reliable it is), the CPU usage
was 40-50% while pulling down 6 Mb.  I've seen a 4501 up to 9-10 Mb on
real internet traffic on 1.2b3, though it was topped out at that.

1.2b7 topped out at 3.2-3.8 Mb at 100% CPU.  

pfSense varied wildly between 3.5-6 Mb with CPU between 70-100%.  

pfSense is currently 5.4-PRERELEASE #1, while 1.2b7 is 5.3-RELEASE #0.
 That might account for the slower-but-steady m0n0wall and faster but
all over the map pfSense.  Might also have nothing to do with it,
since IPFilter and PF, the firewalling and NAT components, are
completely different.

I haven't done any extensive testing without something as
packet-intensive as BitTorrent, but some preliminary testing with only
HTTP downloads shows very similar results.

Maybe this will shed some light on the problem for some of the kernel
hackers out there.

-Chris