Thanks for taking the time for that. I see you're on a mac though and I'm on Windows7. Looking at the history of Recabinet 3, the mac version was always more CPU efficient. If it's 25% more efficient than the windows version, there's the difference. I'm at 256 samples and 96kHz, for what it's worth.
I also wasn't using GR 4 but 5 and didn't load two instances of control room, but one instance of control room pro with different cabinets. I don't know how Control Room Pro functions internally, so I don't know how fair that comparison is. The other test was two parallel instances of Meldaproduction Multiband Convolution with Recabinet IRs. That was an 8% CPU load. To get up to where Recabinet 3.1 is I have to load five in parallel. I'm not sure how fair the comparison is to MMultiband because they go in and out in stereo for every instance.
Are you sure Amplifikation Creme only loads one IR per instance even if you select two different speakers? I was under the impression that the cab sim setup was like a poor man's recabinet: Two cab IRs per instance (but only three cabs to choose from).
Anyway, I'm pretty happy with 20% in 32bit, but I'd be extatic with 10-15%, that's all.