Is it possible that in the future, instead of a dedicated audio processing card, we'll just be able to perform our audio processing on the (GP)GPU?
I've seen a few VSTs for real time processing (convolution reeverbs if I recall correctly) being accelerated with CUDA. I dunno how well they would work on a videogame.
Searched for "cuda vst" in Google and some things turn up. http://www.liquidsonics.com/software_reverberate_le.htm
The latency is not such a problem for audio engineering but becomes problematic for real-time interactive applications.
How much is too much latency?
At least from what I've seen latency is a problem in audio engineering and music production, people prefer to work with DAWs with <10ms latency for maximum responsiveness (specially when dealing with MIDI controllers). 10ms is too much?