I just got a new PC for running Newsbin (not the reason to buy though), and it looks like it has nothing at all to do with disk I/O or number of cores.
My new PC is a Core i7-860, and when I get it in the slow state, there is almost no disk I/O going on, probably because it is very slow to decode.
When I managed to provoke the state for the decoder, I increased the number of connections to 50, which increased my throughput to about 7-8MB/s. The throughput at this point seems to be limited by CPU, since the CPU load I can measure from Newsbin is approximately 1 core 100% loaded. I'm guessing this means that all download threads run in a single CPU thread?
Furthermore, I can see that the wine service takes a fair amount of CPU during download, and increases a bit during decode. First part of decode is relatively fast, and hits a load of 150% for the Newsbin process. But after going to slow state, this changes radically: The Newsbin process goes down to about 66% CPU load, but the wine process doubles its CPU load. It looks like Newsbin gets into some state where it does excessive system calls, and since Wine might be a bit slower than Windows to service the calls, this can increase the effect.