I am one of the founding members of the aforementioned http://ridewithgps.com/ and as a result, I am a little biased However, I have found our gradient calculations to be right on par with what my Garmin 305 records. We calculate gradient for uploaded rides (using TCX, GPX or KML) using the logged elevation data. The new route planner will have live gradient calculations while plotting routes as well, and will be released within two weeks as part of our complete redesign. Our elevation data source is accurate up to 1/3 arcsecond (10 meter spacing) in the US and territories, and 1 arcsecond (30 meter spacing) outside the US. It is the SRTM based elevation dataset mentioned by rdtompki.

rdtompki: the NED data is available (and we use it) up to 1/9 arcsecond (3 meter spacing between elevation points) in some locations. It is not 100 meters between each point, thankfully!

I have a little more work to do on elevation accuracy, but I am confident now that ridewithgps is the most accurate out there in regards to elevations, even without the next round of improvements that should be coming out shortly. First priority is to get the new look and planner pushed out, then after that it is back to some heavy math to make the low-resolution elevation data available into a high resolution source.

Here is a (non-functional) sneak peak of the redesign:
http://seegodesign.com/ridewithgps/routes.php
http://seegodesign.com/ridewithgps/profile.php
http://seegodesign.com/ridewithgps/route.php

Here is an example of a ride I went on recently, logged with my 305. Mouse over the histogram to see the instantaneous grade at any point.

http://ridewithgps.com/trips/776