I think we might benefit more from solving the select problem, with 8 separate workers as mentioned before you're still stuck to 1024*8 which isn't that much more, found an interesting discussion about select, why it's used and why it should not be used on windows systems:
http://groups.google.com/group/alt.winsock.programming/browse_thread/thread/a041475f7312f9c5

If select was rewritten you could run 4 worker_processes with 10.000 worker_connections each on their own cpu. Windows can handle this just as easy as Linux folks.