By: juanrga (nospam.delete@this.juanrga.com), August 16, 2014 2:56 am
Room: Moderated Discussions
David Kanter (dkanter.delete@this.realworldtech.com) on August 15, 2014 3:01 pm wrote:
> juanrga (nospam.delete@this.juanrga.com) on August 15, 2014 11:39 am wrote:
> > David Kanter (dkanter.delete@this.realworldtech.com) on August 15, 2014 9:41 am wrote:
> >
> > > I also am skeptical that anyone can match Intel in performance while trailing behind by a node. To do that,
> > > it is necessary to offer a very different product with a
> > > different system architecture. E.g., target workloads
> > > where cache does not help and simply slap down more memory controllers and cores (hint: that's a GPU!).
>
> > The ISA advantage will be greatly reduced in the top-end
> > side of the performance spectrum, but will not vanish.
> > Keller mentioned during Core Day conference that his K12
> > core will have a "bigger engine" than its x86 sister
> > thanks to the advantages of ARMv8 over x86-64, which allows to spend more transistors on compute.
>
> I happen to know the differences between those two designs. I'm not really sure it's
> going to translate into a significant performance delta. My guess is maybe 10%.
Care to explain how you got the 10% figure? It looks a bit low for me. Thanks
> > Another key is that Intel process advantage will be reduced.
> > Those server-class ARM SoCs that I >mentioned will
> > be made on 14/16 nm FinFET. Broadwell-EP and Skylake-EP on 14nm FinFET will not have a full node >advantage.
>
> Actually they will. The foundry 16nm process will have substantially worse density. The minimum metal pitch
> for the foundries is 64nm, vs. 52nm for Intel. That's about a 20% difference, which is quite significant.
And a ~50% density advantage is not "a full node advantage" as I mentioned just above. Intel _traditional_ process advantage has vanished, as others agree,
http://www.electronicsweekly.com/news/components/microprocessors-and-dsps/whats-new-14nm-processors-intel-2014-08/
> I'm also rather skeptical that we'll see 16nm FinFET in production before the end of 2015.
> TSMC won't be in high volume on 20nm till 3Q when the new iPhone comes out. I expect that
> 16nm FinFET will take at least a year, and quite possibly more to hit high volume.
TSMC claims _10nm_ risk production for late 2015. 16nm is being produced now (check above link). ARM server 16nm parts will be available during 2015--2016. Broadwell has been delayed again to late 2015.
> juanrga (nospam.delete@this.juanrga.com) on August 15, 2014 11:39 am wrote:
> > David Kanter (dkanter.delete@this.realworldtech.com) on August 15, 2014 9:41 am wrote:
> >
> > > I also am skeptical that anyone can match Intel in performance while trailing behind by a node. To do that,
> > > it is necessary to offer a very different product with a
> > > different system architecture. E.g., target workloads
> > > where cache does not help and simply slap down more memory controllers and cores (hint: that's a GPU!).
>
> > The ISA advantage will be greatly reduced in the top-end
> > side of the performance spectrum, but will not vanish.
> > Keller mentioned during Core Day conference that his K12
> > core will have a "bigger engine" than its x86 sister
> > thanks to the advantages of ARMv8 over x86-64, which allows to spend more transistors on compute.
>
> I happen to know the differences between those two designs. I'm not really sure it's
> going to translate into a significant performance delta. My guess is maybe 10%.
Care to explain how you got the 10% figure? It looks a bit low for me. Thanks
> > Another key is that Intel process advantage will be reduced.
> > Those server-class ARM SoCs that I >mentioned will
> > be made on 14/16 nm FinFET. Broadwell-EP and Skylake-EP on 14nm FinFET will not have a full node >advantage.
>
> Actually they will. The foundry 16nm process will have substantially worse density. The minimum metal pitch
> for the foundries is 64nm, vs. 52nm for Intel. That's about a 20% difference, which is quite significant.
And a ~50% density advantage is not "a full node advantage" as I mentioned just above. Intel _traditional_ process advantage has vanished, as others agree,
http://www.electronicsweekly.com/news/components/microprocessors-and-dsps/whats-new-14nm-processors-intel-2014-08/
> I'm also rather skeptical that we'll see 16nm FinFET in production before the end of 2015.
> TSMC won't be in high volume on 20nm till 3Q when the new iPhone comes out. I expect that
> 16nm FinFET will take at least a year, and quite possibly more to hit high volume.
TSMC claims _10nm_ risk production for late 2015. 16nm is being produced now (check above link). ARM server 16nm parts will be available during 2015--2016. Broadwell has been delayed again to late 2015.