Intel Atoms are utterly useless for that. I'd wager they are useless for most anything that isn't memcpy-bound, so I wonder what makes them put 16 of them into one package.
Fun fact: I was running Blender in non-GPU mode on far lower hardware than that almost 15 years ago.
I'd rather run PFSense, but it's not worth an extra $200+ for just my home use.
We have a c2758 doing 12mpps l3 routing (full BGP table) with 50 ACLs on one of these: https://netgate.com/products/xg-2758-1u.html
The c3000 prototypes are due in a couple weeks.