Dummond D. Slow (mental.delete@this.protozoa.us) on November 16, 2020 4:06 pm wrote:
> Will (anon.delete@this.anon.anon) on November 16, 2020 3:30 pm wrote:
> > Maynard Handley (name99.delete@this.name99.org) on November 16, 2020 2:58 pm wrote:
> > > Maynard Handley (name99.delete@this.name99.org) on November 16, 2020 1:45 pm wrote:
> > > > RType (Bigly.delete.delete@this.this.orangeface.com) on November 16, 2020 1:06 pm wrote:
> > > > > Maynard Handley (name99.delete@this.name99.org) on November 16, 2020 12:46 pm wrote:
> > > > > > Against GeForce GTX 1050 Ti and Radeon RX 560.
> > > > > > Not exactly state of the art, but not exactly ridiculous.
> > > > > >
> > > > > > As far as I can tell, these cards are around let's say 75W.
> > > > > > The state of the art is now about twice their performance at about 85W.
> > > > > > So bottom line is Apple's iGPU at, what, 10W? is half the current state of the
> > > > > > art for "mid-range, not insane power". Would that be a reasonable summary?
> > > > > >
> > > > > > My guess is that next year (beginning of Q2?) we get an
> > > > > > 8-large core M1X, for the iMac and MBP/mini pro, with
> > > > > > double the GPU resources, and so a credible match for the (reasonable power level) state of the art at ~20W.
> > > > > >
> > > > > > https://www.macrumors.com/2020/11/16/m1-beats-geforce-gtx-1050-ti-and-radeon-rx-560/
> > > > > >
> > > > > >
> > > > >
> > > > > Polaris (rx 560) is very old now though.
> > > > >
> > > >
> > > > Old. Not very old!
> > > > Like I said, you can at least eyeball the newer cards here:
> > > > https://en.wikipedia.org/wiki/Radeon_RX_5000_series#Desktop
> > > > I looked at the 85W desktop model as a reasonable successor.
> > > >
> > > > You can see the Apple capabilities here:
> > > > https://images.anandtech.com/doci/16233/2020-11-10%2019_12_29.jpg
> > > >
> > > > Single TFlops and pixel rate are comparable, texture slightly behind. But, just like CPU,
> > > > these numbers conceal as much as they tell us because so much depends on the HW algorithms
> > > > that utilize them. ATI and nV have the advantage of longer experience in those algorithms,
> > > > Apple has the advantage of starting from a cleaner slate and (apparently) very little NIH
> > > > and a willingness to change large parts of the design substantially every year.
> > > >
> > > > I guess over the next week many more comparisons of this sort across a range of functionality will appear.
> > > >
> > >
> > > https://twitter.com/mnloona48_/status/1328427807987302400
> > > 1500 ST
> > > 7500 MT
> > > MBP (so active cooling)
> > >
> > > comparison page:
> > > https://www.cpu-monkey.com/en/cpu_benchmark-cinebench_r23_single_core-15
> > > (I'm not sure how to interpret that. Does HT on each device mean HT
> > > was used? As far as I can tell it does, giving about a 25% boost.
> > > In which case only fair would be to allow each large Apple core to get a small core to help it :-) )
> > >
> > > More seriously the MT page is here
> > > https://www.cpu-monkey.com/en/cpu_benchmark-cinebench_r23_multi_core-16
> > > and M1 slots in with the 6 core HT devices. Which is nice enough,
> > > but more remarkable is just how well M1 seems to scale.
> > > MT is giving us essentially 5x ST, so each small core is giving us the full one quarter.
> > > The x86 devices don't seem to scale as well, getting about 5x worth of performance for their 6 (core+SMT)'s.
> > >
> > > As always, one result, preliminary, blah blah. We'll probably (slightly) better results
> > > and (somewhat) worse results, especially on passive cooled MBA over the next few days.
> > > Also still waiting on a VirtualApple result.
> >
> > Computerbase.de has a good list as well.
> >
> > https://www.computerbase.de/2020-11/cinebench-r23-community-benchmarks/
> >
> > One notable result is the 6/6 4600u which gets 5.7x MT speedup. There are slower
> > results for the 4600u but this one is the only one with a screenshot.
>
> That is not scaling at all, that is artifact of how x86 processors run at much
> higher clock in single-thread boost mode compared to all-core (all-thread rather)
> load. The difference is going to be massive for an U-line processor.
>
> If you want to evaluate "scaling", you gotta lock the frequency to a constant value.

Well it depends what you're looking for as to what "scaling" means.
If you know a priori that the problem is trivially parallelizable, the only interesting scaling number is what happens when a constraint kicks in - either thermal or (in a particularly badly unbalanced system) memory bandwidth.

Thermal scaling is especially interesting precisely because the second pillar of the "ARM will conquer the world over the next few years" is precisely that it gives you this much better "practical" scaling in most of the cases that matter, most obviously data warehouses.