you can try to run some benchmarks ( for example Swingbench, very easy to setup) and compare results.
I think you don't find big difference.
For nonRAC databases you have both options PV and HVM, for RAC you have only PV to be supported.
With such a big number of vCPUs your VM will have quite a lot of RAM and can support large number of users.
You can consider the use of large pages, which shloud be supported for HVM only (probably supported for PV in the future or unsupporded now) and should work out of box.
You can easy test the advantage of large pages using prepage sga parameter with large sga and many users, I think this emulates the worst scenario :-)
We successfully run hard partitioned PV guests (12 vCPUs and 72GB, 6 vCPUs and 54GB) on server with one 10GB CNA adapter. (OEL 5 UEK2)
The VMs are capable to do constantly 15k IOPS and 400 MBytes/s to disk array and utilize network bandwidth. (SAN physical disks)
Preproduction tests show almost theoretical throuput (disk and network).
Testing CPU-bound parallel execution - the scalability was linear ( parallel degree <= vCPU count) and almost didn't drop (parallel degree <= twice vCPU count).
Can you, please, point some links to the "known PVM limitations with CPU scalability" ?
I agree with user459387. I basically use PVM on all my hosts and have never had an issue. My prod RAC is a 2 node PV with 12 vcpu and 100 GB Ram each. I have seen as much as 40,000 IOPS over FC. I run OEL 5U5 64 bit UEK2 guests. I understand Oracle VM to primarily be paravirtual focused product with many enhanced to improve paravirtual guests performance.
Hugepages is supported in Oracle VM 3.2.4 and later with OEL 5.8 and higher and OEl 6.3 and higher. There is a really detailed MOS article on using this with PV guests.