When you are creating your own images*, it is generally a good idea to work with dimensions fairly close to those at which your final result will be presented. As a general rule, no more than twice the size (in most situations, I'd advise using the same dimensions).
If you are working at 3300x2300 and later downsizing to 640x400 then this means that for every 25 pixels in the original work, you only get one pixel in the final. The effect may not be what you want. For example, the following graphic shows a bright blue line being downsized to a single, light-blue pixel -- is that what you'd want? Maybe. Or maybe you'd rather the pixel be either the original white or original blue. The problem is that no scaling algorithm can figure out your intent or what information in the original is most meaningful to you (and you are effectively eliminating 95% of the color information by scaling down by a factor of 5).
Attachment:
downscale-by-5.png [ 723 Bytes | Viewed 1896 times ]
* When working with pre-existing content such as photographs, it is usually best to work at the original size until the final stage and then scale the work down to presentation size.