> Linus Torvalds (torvalds.delete@this.linux-foundation.org) on August 13, 2018 2:12 pm wrote:
> wumpus (lost.delete@this.in.a.cave) on August 13, 2018 9:32 am wrote:
> > Per Hesselgren (perhesselgren.delete@this.yahoo.se) on August 13, 2018 8:50 am wrote:
> > > Threadripper on Anandtech
> >
> > Looks like coherency isn't an issue for 32 cores, but obviously not everything scales perfectly.
>
> Even loads that should scale really well (like the chromium build) seem to have hit the
> point of negative scaling here, with a 32c/64t 2990WX doing worse than a 16c/32t 2950X.
>
> I do suspect it's mostly the load that doesn't scale any more at that point, with the negative scaling coming
> from more memory contention and perhaps half of the cores being further away from the memory controller.
>
> Somebody who knows the Chromium build better might be able to tell how far they'd expect
> the build to scale, but even for simple things like a parallel build, you often hit
> serialization points (eg linking phases after you've built all your library files).
> And writing Makefiles that expose all the parallelism isn't trivial either.
>
> But some of that may be "nobody spent any effort to scale past X cores because they are so
> rare, and build farms have their own scaling issues where IO may be the bottleneck".
>
> So some of the scalability issues may just be "people haven't even
> tried to get that much parallelism". Others may be more fundamental.
>
> Linus

Here are some reviews from Phoronix:

<a href="http://">2990WX on Linux

<a href="http://">2990WX on Windows 10 Pro vs 4 Linux distros stock

Essentially it seems that Linux distros don't have a scaling problem to 32/64 cores/threads. Windows 10 Pro is the problem.

Pete