I've got a patch for this.  It looks like they're sending an abnormal
interrupt every time any transmit completes.  I've hacked in a new
tulip type, to try to mask those interrupts, but that didn't help,
so I *also* just always mask them out if cpu_is_vpc is non-zero.

My patches now distinguish between VPC3 and VPC4, and most of the patches
only change anything on VPC3 now.  I'll be submitting the "current" version
later today, if all goes well, probably via send-pr.

-s