I think as computers fail to get faster, we won't be able to improve performance without making application specific logical units, however, we may still be able to make it general purpose.
What I'm most interested in is what the trends will be in how those specialized circuits are developed.
Circuit design has been marching steadily along for decades, and between the tools available to allow non-specialists to go from abstract hardware description to image-ready chip specification and the ever-cheapening, ever-broadening capabilities of custom chip fab, are we approaching an era where fabricating mid-range specialized circuits is on par with the difficulty of writing and compiling software, only a bit slower?
I'm waiting for automatic code optimization. Write the code going from an input to an output however you like, show the computer your input and output, have it rewrite the most optimal (and probably very weird looking) code possible to go from your supplied input to your supplied output, and reap the efficiency rewards to mastering a stack of arcane low level code reference books without spending the thousands of hours of your precious and limited free time to do so. That would be a huge leap imo.
