Rayla (rayla.delete@this.example.com) on September 7, 2021 7:51 am wrote:
> Daniel B (fejenagy.delete@this.gmail.com) on September 7, 2021 2:31 am wrote:
> > anon2 (anon.delete@this.anon.com) on September 6, 2021 3:36 pm wrote:
> > > Result here
> > >
> > > Not a bad result compared with the recent quite equivalent HPE Superdome
> > >
> > > About 50% faster per socket, but about 50% slower per thread (and 2x the number of threads).
> > > I think POWER has always done quite well on this benchmark for some reason (maybe it likes thread
> > > level parallelism?) so it remains to be seen how this translates to other performance.
> > >
> > > The POWER10 is nearly as fast as a POWER9 with 2x more sockets though, so comparing POWER vs POWER
> > > there is still significant gain. Thread strength up about 50% and threads per socket up about 50%.
> >
> > Technically, I feel you are underselling Power10 quite a bit. 1.42x socket throughput is married by 2.66x
> > core performance. With that kind of gap, I expect the Power10 system should be comfortably ahead still
> > when set to SMT4 or SMT2 even (with lower throughput advantage), should it come to a competitive stand-off
> > with x86 systems. Which is probably not really the key selling point. No amount of technical superiority
> > has been able to reverse the erosion of Power systems sales, so it stands to reason to assume per thread
> > performance, particularly when improved generation over generation, will not be an issue.
> >
> > More enlightened folks will be able to analyse the performance
> > characteristics better, but what I understand
> > is that IBM has always equipped Power systems with much higher memory, interconnect and IO bandwidth than
> > comparable x86 systems (and designed the cores/uncore with those assumption) because the high ASP of Power
> > systems can absorb the added manufacturing costs, and the targeted applications benefit greatly.
>
> That "core performance" claim is kind of exaggerated by the fact that what IBM calls a Power10 "SMT8 core"
> is a pair of "SMT4 cores" sharing almost nothing. IBM has previously done the same thing with Power9 (and
> Fujitsu did with the SPARC64 XII) and I tend to think it really obfuscates actual core performance.
>
> By the way, SPEC is up now too.
>
> https://spec.org/cpu2017/results/res2021q3/cpu2017-20210814-28679.html
>
> 2-socket ICL-SP for comparison:
>
> https://spec.org/cpu2017/results/res2021q3/cpu2017-20210720-28374.html

You can actually accurately estimate the commercial performance when scaling back from smt8 to smt4, smt2 or st.

IBM gives the numbers in their performance report

https://www.ibm.com/downloads/cas/K90RQOW8

No power10 in there yet but power9 is there.

For the fully loaded e980
St 1719
Smt2 2922
Smt4 4032
Smt8 5081