my inner geek writes:
> Yes. Distributed HDTV wavelet compressed video.
The problem of video compression is primarily numerics-limited, there is nothing an OS can change here. The latency and bandwidth of the Internet is so bad that the OS overhead is invisible. As to parallel DSP boxes, for that particular application you don't need an OS at all, or a very skeletal one. There is a number of minimal DSP OSSes out there, particularly for TI's chips. It is not very hard writing these.
Btw, a related idea: a few days ago some friends of mine asked me whether a modern video accelerator could be useful for neural DSP (as you know, you could use several in a PCI system). I am aware that next-generation video chips will feature VLIW DSPs with multiple GOPs, but do _current_ video accelerators qualify? I had the impression that these things were pretty special-purpose...