Currently, there are tools that allow you to take advantage of CUDA without a single line of CUDA code ( ex. caffe2, opanACC … ). What is the current state of these technologies compared to hand-tuned CUDA programs? And is it possible, within a near foreseeable future, that such tools become so effective that we no longer need people to use the CUDA runtime / driver API directly ( effectively rendering cuda engineers useless )?
Since there are so many applications for iphone available in the App Store, will iphone application developers lose their job in the foreseeable future?
I think you can answer that yourself.
This is essentially what you’re asking here, except it’s about CUDA.
CUDA developers will always find something to develop, even if it’s just maintaining the existing code base and making it keep up with new hardware generations. But chances are, that there are new and exciting CUDA applications to develop that no one has thought of before.
I was looking more for numbers that would show rather or not we should hand-tune CUDA or just let the APIs do the work… Do you have any of those ( or even just by your general experience with both )?