> Real speed of G5 970 CPU on the same frequency is lower than G4 and G3

As I wrote, for a fair comparison it's important that the binaries executed by the PPC970 do not contain instructions that the G5 lacks, else what you're actually measuring is how fast the G5 can emulate certain portions of the G3/G4. Ideally, binaries come in at least two fashions, one for pre-G5 (which could be two because of AltiVec) and one for G5. Except for E-UAE, I'm not aware of G5-specific binaries for MorphOS (and I doubt there are G5-specific code paths in the binaries). So considering that the very same MorphOS binaries do run on G2 (Efika) to G5 (as well as several Book III-E microarchitectures), what the G5 is actually running is generic PPC code not adapted to the peculiarities of the PPC970 instructions and microarchitecture, including missing instructions it has to emulate. As a result, I'm sure MorphOS programs on G5 could be even faster with binaries compiled for 32-bit PPC970.
I wrote about these things 3 years ago in another thread (comments #2 and #4).

> I cannot say for sure what is exact value of G5 MIPS

Just to make sure we're not talking past each other: What l'm referring to is not just MIPS (which is really meaningless) but Dhrystone 2.1 MIPS (DMIPS). Dhrystone 2.1 is a specific version of a specific benchmark algorithm.

> If you wanted to take the time and found a real source from IBM, I will be be glad.

At your service:

https://bunniestudios.com/blog/images/efuse3.pdf (p. 1, 6)
http://datasheet.digchip.com/205/205-00367-0-970.pdf (p. 13, 14)