Thu (Anon.delete@this.anon.com) on August 17, 2020 3:27 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on August 17, 2020 2:12 am wrote:
> > Thu (anon.delete@this.anon.com) on August 16, 2020 11:56 pm wrote:
> > > Crystal S. Diamond (cdiamond.delete@this.diamondgirls.com) on August 16, 2020 10:20 pm wrote:
> > > > Here it is boys...
> > > >
> > > > https://www.hardwareluxx.de/index.php/news/hardware/prozessoren/53864-ibm-power10-bietet-30-kerne-mit-smt8-pcie-5-und-omi-memory.amp.html?__twitter_impression=true
> > >
> > > Does look very impressive. With 3X the performance per socket
> > > using the same energy as POWER9 - it should easily
> > > beat Intel and AMD at time of release in terms of performance.
> >
> > Beat in which metric?
> > Per system, per chip, per core or per thread?
> > Per core throughput lead would be easy, but that metric is meaningless for most of us.
> > Lead in other metrics would be much harder.
> >
> > State of the art:
> > SpecInt2017 rates (1 chip)
> > AMD EPYC 7H12, 1S/64C/128T, 2.6GHz - 397/361
> > Intel Xeon Platinum 8280L, 1S/28C/56T, 2.7 GHz - 180/171
> >
> > 1-chip POWER9 is not available. There are only two scores:
> > IBM Power E950, 4S/40C/320T, 3.4 GHz - 475/392
> > IBM Power S924 2S/24C/144T, 3.4 GHz - 277/213
> >
> > > Not sure about energy efficiency though.
> > >
> >
> > For me the most interesting would be if IBM is going to publish
> > non-rate benchmarks. It didn't happen for a long time.
> > https://www.spec.org/cpu2006/results/res2010q2/cpu2006-20100426-10753.html
> > 10.5 years ago.
> >
> Just some rough calculations suggests that a power10 socket with 30 cores could get around 480+ in
> specint if you take the 3x performance they talking about is comparing the highest end power10 Vs power9.
> Highest end power9 is clocked at 4ghz - quite a lot higher than the 3.4ghz submitted figure.
>
> But I doubt IBM is overly concerned with specint - they would
> target OLTP workloads which should excel in these machines.
>
> I doubt ibm will ever publish any non rate benchmarks - it is guaranteed to be lower than competition.
> Power10 is only increasing single thread performance by 20%. - so that would hardly catch up.
>
>

I think you're essentially right. P10 looks like a combination of "it beats Epyc2 on throughput" and "it has a lot of interesting I/O capabilities."

Direct-attached memory is looking pretty dead, which is a pity.