As of today the support for NUMA topology is somewhat rudimentary. 
Roman did some of the basic implementation 
https://github.com/kubevirt/user-guide/pull/457/files

But this has some nondeterministic behavior in the sense that kubevirt can not provide specific cores from specific NUMA nodes. This "might work" for the single big VM scenario, but is not sufficient to support multiple VMs. We need to be able to reliably acquire certain cores from specific NUMA nodes.  

What we need is a way to define how many NUMA nodes (and sockets), cores and threads a VM should have and the allocated resources should live 1:1 on the underlying hardware, e.g. the CPU cores from the same NUMA nodes bare metal and virtual and same for the memory attached to it. 

The sizes we should plan for are 1,2,4,... NUMA nodes per VM and if we can plan for sub NUMA node entities as well (e.g. "half" a CPU socket), that would be great.

The current implementation is specified as follows:

    spec:
      domain:
        cpu:
          cores: 10
          sockets: 4
          threads: 2   
          dedicatedCpuPlacement: true
          isolateEmulatorThread: true
          model: host-passthrough
          numa:
            guestMappingPassthrough: {}


Either this is being enhanced by the numa node count attribute, e.g.:

    spec:
      domain:
        cpu:
          cores: 10
          sockets: 4
          threads: 2   
          dedicatedCpuPlacement: true
          isolateEmulatorThread: true
          model: host-passthrough
          numa:
            guestMappingPassthrough: {}
            nodes: 4

or it's implicitly derived from the socket count when guestMappingPassthrough is set. I guess that's up for discussion since there might be platforms where there is no 1:1 mapping between socket and NUMA node.