Wrong application of GEOMEAN - use just arithmetic mean!

[Kocicak]

I noticed the recent trend of using geometric mean instead of arithmetic mean when averaging results of benchmarks. Geometric mean is a very specific tool, intended primarilly for averaging growth rates in time series.

There is NO REASON AT ALL to use it for "general averaging"!

 

[Sgt. York]

That's just, like, your opinion, man.

 

[Kocicak]

That is not my opinion, that is like using wrong statistical method. The results are either meaningless totally or they do not bring anything better than the proper method.

 

[serpretetsky]

Can you link to a specific example for reference?

 

[TheELF]

Can you link to a specific example for reference?

Geomean gives less importance to results with high differences.

AMD Ryzen 9 3900X SMT Linux Performance Benchmarks - Phoronix

www.phoronix.com

Looking at geomean alone is ok for users that only care about general usage,like here you see that SMT on vs off is about the same performance in the grand scheme of things.

But when looking at all results you can see that more apps respond better to SMT on, so even a general usage user will have better chances of running something that will benefit more from SMT.

Obviously if you are a user that primarily uses one of those 3-4 big outliers then geomean will screw you over big time.
So as you can see it all depends on what you gonna do with your system.

 

[serpretetsky]

Geomean gives less importance to results with high differences.

AMD Ryzen 9 3900X SMT Linux Performance Benchmarks - Phoronix

www.phoronix.com

Looking at geomean alone is ok for users that only care about general usage,like here you see that SMT on vs off is about the same performance in the grand scheme of things.

But when looking at all results you can see that more apps respond better to SMT on, so even a general usage user will have better chances of running something that will benefit more from SMT.

Sorry, I meant a link where the arithmetic mean clearly makes more sense. Taking the arithmetic mean of scores/times of completely different tests doesn't necessarily give you a better picture of the performance of the cpu, but it will skew the score of the CPU towards tests that give very high values.