I just got an idea that it might be interesting to combine mod_dims and GPU accelerated image manipulation libraries like NVidia's NPP. Apparently nobody looked into this so far, even though a google search will show a lot of GPU related image processing stuff, NVidia also provides a nice intro to GPU image processing.
The idea is that if it would be possible to service scaled images as fast as unscaled images then it would be possible to skip all the solutions that deal with the caching of scaled images and just scale the images always on demand. This depends on the ability to service scaled images really fast, for example with the help of some Tesla boards for the image scalers. It would be an interesting comparison to compare the cost of sufficiently available storage for caching plus the cost for elastic on-demand scaling with the cost of permanent on-demand scaling with GPU acceleration.
The advantage of the latter would be, if the costs are in the same region, that it would be a way to be independent of external resources and keep everything in our own data centre.
I am therefore looking for people who like this idea and would like to join in a Proof-Of-Concept implementation of this idea (first with the help of commodity NVidia GPUs in a desktop) to get a first taste of the power of GPU accelerated image scaling. Please contact me if you want to join me in this attempt.
Update: See also GPU Computing - Know When to Use It.