Some basic PowerVM-LX86 benchmarks, side-by-side (left PowerVM-LX86, right native ppc64) with bogomips from
Contribute to aib/bogomips development by creating an account on GitHub.
github.com
I compiled it under i586 OpenSuse 11.2 to have an SLES-11 compatible glibc (glibc-2.11.1-0.17.4 for SLES11 and glibc-2.10.1-10.4 for OpenSUSE 11.2), as compiling it under Tumbleweed live i586 (20230106) results in an incompatible binary since it requires glibc 2.34. Cross-compiling (not possible under Tumbleweed ppc64, i586 cross-compiling is not supported anymore it seems at least under ppc64) under debian ppc64 or ubuntu x86-64 also resulted in incompatible binaries. The only difference in the compilation was that I had to use
under i586 OpenSuse 11.2.
The bottom line is that the performance looks very weak,
5.5% efficiency at best.
Geekbench2 seemed to indicate slightly better performance,
at around 15% comparing with native osx ppc64 benchmarks, e.g.
Benchmark results for a Power Mac G5 (Late 2005) with a PowerPC G5 (970MP) processor.
browser.geekbench.com
I have noticed that under Tumbleweed generally speaking (both ppc64 and i586), cpu performance is generally weaker than under debian (notably via hardinfo) so I've copied all the PowerVM-LX86 directories (/opt/powervm-lx86 and /opt/at05) to debian ppc64, I will see if I can get it to work there and if it's any better.
It could be that powervm-lx86 does somehow make use of the MSR endianness switch bit under POWER5/6/7 (absent on the G5) but that there is a fallback mode for the G5 albeit much less efficient as there could be an additional byte-swapping step in the dynamic instruction translation, but maybe not.