The use cases are much different. Once you get outside of the standard parallel sort/scan/reduce/map functions you can find in Thrust[1], you get into algorithms that have an entire enormous book[2] explaining the idea behind them.
That gap is closing slowly as a lot of old algorithms are being remixed into highly parallel versions. This is noticeable in string pattern matching applications, driven by the need to analyze tons of DNA.
More people would benefit from knowing how exactly to code efficiently on the GPU before jumping right in to writing algorithms on it; it's easy to lose multiples of throughput if you don't know what you're doing. As always it's becoming easier for programmers though; if you're new to the idea of GPU programming I really recommend checking out Microsoft's C++ AMP[3], which shipped with VS 2012. Removes a lot of the headaches of programming with CUDA and OpenCL. I've had a chance to play around with it, and it's excellent.
That gap is closing slowly as a lot of old algorithms are being remixed into highly parallel versions. This is noticeable in string pattern matching applications, driven by the need to analyze tons of DNA.
More people would benefit from knowing how exactly to code efficiently on the GPU before jumping right in to writing algorithms on it; it's easy to lose multiples of throughput if you don't know what you're doing. As always it's becoming easier for programmers though; if you're new to the idea of GPU programming I really recommend checking out Microsoft's C++ AMP[3], which shipped with VS 2012. Removes a lot of the headaches of programming with CUDA and OpenCL. I've had a chance to play around with it, and it's excellent.
[1] http://thrust.github.com/
[2] http://books.google.ca/books/about/Computational_Electrodyna...
[3] http://msdn.microsoft.com/en-us/library/hh265137.aspx