> The '*Local* Communication bandwidths' numbers don't looks rigth compared
> to Simon's results for a ds20/OSF1. I can't believe NetBSD is that bad :)

The EV6 can have 2 memory operations in progress at the same time.  
The performance gain is dramatic when you can take advantage of that. 

Are the local communication numbers testing the compiler or bcopy 
and friends?  Have either been tuned to know about EV6?