[See edit 2 below: problem solved, but some mystery remains]
I’ve been swapping CPUs quite a bit lately, going from one 3970X to a 3990X, back to another 3970X, etc. The system (Windows 10 Pro) is always the same and is a rather fresh install.
I’m now back to a 3970X that I used in the past, and surprisingly performances are significantly lower than before, on all benchmarks that I’ve been tested (Cinebench R20, V-Ray Benchmark 4.10 and a few others). For instance, on Cinebench R20 I’m getting around 16000-16300 points where I used to get 16700-16900. On another rendering benchmark, I’m getting 1 min 09 instead of 56 seconds.
XMP profile is set and memory frequency is correct (3466 MT/s in my case), SMT is enabled, ambient temperature is the same as in my other tests (20/21 °C). What other causes am I not thinking about?
I’m kind of suspecting a less lucky application of thermal paste this time (Noctua NT H2), maybe a tad more of it than usual. Could this be it?
Edit 1: I can actually see that at the beginning of the benchmark, a large number of cores boost to 4.2 or 4.3 GHz, but quite quickly they all come down to 3.8 GHz and stay there for the remaining of the benchmark.
Edit 2: On a hunch, I went to the BIOS and disabled SVM (virtualization). It turns out that SVM was the culprit: benchmarks are back to expected performances. I remember that at some point I enabled Windows 10’s sandbox feature and I suspect that enabled some virtualization features causing additional overhead on the host, even when not running any VM. A cursory look on Google seems to confirm this. I then rebooted once more, enabled SVM again and it seems that performances are still holding. What’s happening?