By: Wilco (Wilco.Dijkstra.delete@this.ntlworld.com), June 3, 2013 9:52 am
Room: Moderated Discussions
Exophase (exophase.delete@this.gmail.com) on June 3, 2013 9:31 am wrote:
> anon (anon.delete@this.anon.com) on June 3, 2013 8:54 am wrote:
> >
> > > Interesting slide from ARM at Computex shows 28nm A15 perf than
> > > Silvermont at 22nm FF. Dispels Intel's Marketing FUD?
> >
> > http://www.pcworld.com/article/2040582/arm-claims-processor-superiority-over-intels-silvermont.html
> >
>
> I take Intel's marketing with a lot of salt but at least they have realistic access to real Cortex-A15
> hardware. The same can't be said for ARM and Silvermont.
On the other hand the only A15 implementation available until a few weeks ago (Exynos 5250) is unlikely to be indicative of all A15 implementations.
>Furthermore, most of Intel's claims come
> from comparing vs Saltwell, where they're more likely to do a fair test running the same binaries.
> ARM has nothing to go on but Intel's numbers which completely contradict their claims.
>
> The power vs efficiency graph is completely insane if you ignore Silvermont and only look at Saltwell. First
> of all, how does Cortex-A7 even go up to Saltwell's performance when no one is clocking it beyond 1.2GHz vs
> up to 2GHz for Saltwell (in relevant plaforms, higher for desktops) and A7 is going to typically have considerably
> less perf/MHz than Saltwell. Yet ARM says A7 can deliver that same performance with 5x less power?
If you compare quad A7 vs dual Atom then yes that appears quite possible. Note the A7 does currently clock up to 1.4GHz and is only a little slower than A9 due to a much better memory system.
> For reference, ~182mW @ 1.2GHz for Cortex-A7 on Samsung 28nm HKMG (http://beyond3d.com/showpost.php?p=1732331&postcount=530).
> Intel says for Saltwell 175mW for 600MHz, 500mW for 1.3GHz, and 750mW for 1.6GHz on what should be a competitive
> 32nm HKMG process (http://www.anandtech.com/show/5365/intels-medfield-atom-z2460-arrive-for-smartphones/2).
> Real power consumption numbers show that these estimates are fair.
According to Samsungs ISSC graph, a quad A7 at 1.2GHz uses ~500mW, or just ~125mW per core.
> Let's say A7 @ 1.2GHz is similar to Saltwell @ 1GHz, at least where the software is fairly level for
> both. A cubic fit shows that to be around 320mW. So Cortex-A7 is definitely more efficient here but
> nowhere close to 5x, and comes at a price of not being able to hit nearly as high peak performance.
Unless you use extra cores of course. Assuming a 1.2GHz A7 is about as fast as a 1GHz Atom, a quad A7 at 1.2GHz should be able to match dual Atom with hyperthreading at 2GHz. Extrapolating your Atom power figures, Atom would use ~1170mW per core at 2GHz. So it would be 2343mW vs 500mW, close to a factor of 5 as claimed by the ARM slide.
> I won't even touch on the Cortex-A15 part of the graph, which is probably more ridiculous.
I don't see anything obviously incorrect there either. It's pretty similar to the graphs Samsung showed at ISSCC.
> I like ARM and I want them to succeed and be competitive, but this sort of marketing is just embarrassing..
Well it isn't like Intel isn't making big claims about Silvermont like they were about the original Atom. I'd like to see the full disclosure (clocks, core counts and benchmarks used), but after repeating the 5x result using the best available data for Atom and A7, I believe it may well be 100% correct.
Wilco
> anon (anon.delete@this.anon.com) on June 3, 2013 8:54 am wrote:
> >
> > > Interesting slide from ARM at Computex shows 28nm A15 perf than
> > > Silvermont at 22nm FF. Dispels Intel's Marketing FUD?
> >
> > http://www.pcworld.com/article/2040582/arm-claims-processor-superiority-over-intels-silvermont.html
> >
>
> I take Intel's marketing with a lot of salt but at least they have realistic access to real Cortex-A15
> hardware. The same can't be said for ARM and Silvermont.
On the other hand the only A15 implementation available until a few weeks ago (Exynos 5250) is unlikely to be indicative of all A15 implementations.
>Furthermore, most of Intel's claims come
> from comparing vs Saltwell, where they're more likely to do a fair test running the same binaries.
> ARM has nothing to go on but Intel's numbers which completely contradict their claims.
>
> The power vs efficiency graph is completely insane if you ignore Silvermont and only look at Saltwell. First
> of all, how does Cortex-A7 even go up to Saltwell's performance when no one is clocking it beyond 1.2GHz vs
> up to 2GHz for Saltwell (in relevant plaforms, higher for desktops) and A7 is going to typically have considerably
> less perf/MHz than Saltwell. Yet ARM says A7 can deliver that same performance with 5x less power?
If you compare quad A7 vs dual Atom then yes that appears quite possible. Note the A7 does currently clock up to 1.4GHz and is only a little slower than A9 due to a much better memory system.
> For reference, ~182mW @ 1.2GHz for Cortex-A7 on Samsung 28nm HKMG (http://beyond3d.com/showpost.php?p=1732331&postcount=530).
> Intel says for Saltwell 175mW for 600MHz, 500mW for 1.3GHz, and 750mW for 1.6GHz on what should be a competitive
> 32nm HKMG process (http://www.anandtech.com/show/5365/intels-medfield-atom-z2460-arrive-for-smartphones/2).
> Real power consumption numbers show that these estimates are fair.
According to Samsungs ISSC graph, a quad A7 at 1.2GHz uses ~500mW, or just ~125mW per core.
> Let's say A7 @ 1.2GHz is similar to Saltwell @ 1GHz, at least where the software is fairly level for
> both. A cubic fit shows that to be around 320mW. So Cortex-A7 is definitely more efficient here but
> nowhere close to 5x, and comes at a price of not being able to hit nearly as high peak performance.
Unless you use extra cores of course. Assuming a 1.2GHz A7 is about as fast as a 1GHz Atom, a quad A7 at 1.2GHz should be able to match dual Atom with hyperthreading at 2GHz. Extrapolating your Atom power figures, Atom would use ~1170mW per core at 2GHz. So it would be 2343mW vs 500mW, close to a factor of 5 as claimed by the ARM slide.
> I won't even touch on the Cortex-A15 part of the graph, which is probably more ridiculous.
I don't see anything obviously incorrect there either. It's pretty similar to the graphs Samsung showed at ISSCC.
> I like ARM and I want them to succeed and be competitive, but this sort of marketing is just embarrassing..
Well it isn't like Intel isn't making big claims about Silvermont like they were about the original Atom. I'd like to see the full disclosure (clocks, core counts and benchmarks used), but after repeating the 5x result using the best available data for Atom and A7, I believe it may well be 100% correct.
Wilco
