I don't know the exact ins and outs of it so I may be biting off more than I can chew, but here's what I think:
A single CPU is busy doing the following things while in a game:
[*]Sending commands to the graphics card
[*]Calculating AI (single player)
[*]Calculating physics
[*]Calculating general gameplay (not very intensive)
So with all the other stuff on the CPU, the graphics card command sender is being squished like 8 people in the back seat of a car.  Obviously we can't just not process the physics, so they all get their share of time.  So that would make the graphics command processor use about 25% of the CPU time or so.  Well in the games that do get increases, that part of the processing is bottlenecked.  Keep in mind the second CPU is not in use at ALL in this point (for 99% of the single-threaded games).  So with the multithreaded driver, it can get its 25% share on the first CPU plus another 100% on the second CPU.  So it's not bottlenecked anymore... 
It's hard to say because I don't know how intensive sending graphics commands is, but the results seem to support it, unless the increases are coming from something else new in the driver which is likely too.  Remember that thread about NVIDIA holding back 10% performance?  It could be that too.  Increases will vary greatly per game though.