it may be better now but our initial testing, years ago, showed a pretty significant perf impact for rtp streaming.
The limiting factor for MBG isn't cpu, it's network interrupt load and the extra path through the hypervisor, the virtual "network", and the virtual network drivers in the guest VM do have a big impact on the max interrupts/sec that can be processed and that means it limits the number of active calls that can be handled simultaneously.
that's all.