I benchmarked a system both ways - exact same hardware, exact same pools, same combination of SE11+Napp-it. Compared bare-metal vs ESXi 4.0/all-in one. There was no materiel difference in performance using CIFS/SMB to windows clients. There was a slight difference when using an advanced NFS client on the windows side (OpenText/Hummingbird), but not enough to care - less than 10%. This was true even over a 10Gbe physical link - no materiel difference in performance for reads or writes. There was no materiel difference between running it bare metal vs all-in-one under ESXi.
If you've read other threads of mine here you'd see that I ended up leaving it running bare-metal. But the only reason was due to a problem with restart stability of SE11 under ESXi 4.0. These problems are fixed under ESXi 5.0 and should no longer be a reason to avoid all-in-one. I'm leaving it as is because I am in "production" for my home network and have no motivation to change, but based on what I know today, if I were building it again I believe I would choose the all-in-one approach. It would have saved me the cost of deploying a separate server for ESXi-based VMs.