By: Michael S (already5chosen.delete@this.yahoo.com),
Room: Moderated Discussions
anon (anon.delete@this.anon.com) on February 20, 2019 8:27 pm wrote:
> Linus Torvalds (torvalds.delete@this.linux-foundation.org) on February 20, 2019 9:36 am wrote:
> > nobody in particular (nobody.delete@this.nowhe.re) on February 20, 2019 7:35 am wrote:
> > > Write-up at Anandtech
> > >
> > > The stated SPECint numbers are superb. I'm looking forward
> > > to spec2017, but this is a very, very promising start.
> >
> > We'll see if it actually delivers, but ARM is certainly looking a whole lot better than it used to.
> >
> > They've been nicely strengthening their memory model, to the point that these days it's actually
> > one of the better ones. They even seem to have decided that I$ coherency matters. Good for them.
> > Doing cache coherency in software is crazy, and people who think it's a good idea don't understand
> > the complexities of reality. Maybe some day it will even become architectural.
> >
> > And I have to say, I like the direction ARM is going with vector math a lot more than the
> > AVX512 that Intel is pushing. I don't know how well it works in practice, but the whole "let's
> > try to do something that works for different vector lengths" is laudable. I'm quite tired
> > of the model where Intel introduces yet another incompatible model every few years.
> >
> > I still will hold judgement until we actually see widely available hardware that people
> > actually can use for development and deployment. I've just seen too many promises and "released"
> > hardware that never went anywhere and nobody really had reasonably available.
> >
> > In particular, I do wish they didn't push the "hyperscale" design so much. Maybe they got the
> > scaling working, but honestly, I doubt it. It just takes time and effort and learning. Don't
> > try to jump to 64-128 core targets until you've had a few years of just plain getting something
> > like a "simple" 8-core one right. Which they haven't actually demonstrated, so far.
>
> Why? They have no market for an 8 core design. They would have to not just match but beat
> x86 in single threaded performance for there to be any reason to use such a thing.
Linus is the ultimate unixoid. I paid attention that even less devoted unixoids are high on native development.
For me, as one that drinks and breaths cross-development all his professional life, it sounds strange, but this mindset is not rare at all.
> Linus Torvalds (torvalds.delete@this.linux-foundation.org) on February 20, 2019 9:36 am wrote:
> > nobody in particular (nobody.delete@this.nowhe.re) on February 20, 2019 7:35 am wrote:
> > > Write-up at Anandtech
> > >
> > > The stated SPECint numbers are superb. I'm looking forward
> > > to spec2017, but this is a very, very promising start.
> >
> > We'll see if it actually delivers, but ARM is certainly looking a whole lot better than it used to.
> >
> > They've been nicely strengthening their memory model, to the point that these days it's actually
> > one of the better ones. They even seem to have decided that I$ coherency matters. Good for them.
> > Doing cache coherency in software is crazy, and people who think it's a good idea don't understand
> > the complexities of reality. Maybe some day it will even become architectural.
> >
> > And I have to say, I like the direction ARM is going with vector math a lot more than the
> > AVX512 that Intel is pushing. I don't know how well it works in practice, but the whole "let's
> > try to do something that works for different vector lengths" is laudable. I'm quite tired
> > of the model where Intel introduces yet another incompatible model every few years.
> >
> > I still will hold judgement until we actually see widely available hardware that people
> > actually can use for development and deployment. I've just seen too many promises and "released"
> > hardware that never went anywhere and nobody really had reasonably available.
> >
> > In particular, I do wish they didn't push the "hyperscale" design so much. Maybe they got the
> > scaling working, but honestly, I doubt it. It just takes time and effort and learning. Don't
> > try to jump to 64-128 core targets until you've had a few years of just plain getting something
> > like a "simple" 8-core one right. Which they haven't actually demonstrated, so far.
>
> Why? They have no market for an 8 core design. They would have to not just match but beat
> x86 in single threaded performance for there to be any reason to use such a thing.
Linus is the ultimate unixoid. I paid attention that even less devoted unixoids are high on native development.
For me, as one that drinks and breaths cross-development all his professional life, it sounds strange, but this mindset is not rare at all.
