By: Wilco (Wilco.Dijkstra.delete@this.ntlworld.com), June 3, 2013 1:47 pm
Room: Moderated Discussions
Exophase (exophase.delete@this.gmail.com) on June 3, 2013 12:02 pm wrote:
> Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on June 3, 2013 10:52 am wrote:
> > 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.
>
> No, but it's better than the Silvermont hardware ARM has access to, which is nothing.
True. But Intel's claims would not be correct if they are based just on the 5250 and assume no further improvement (tuning, big.Little, 20nm process change etc).
> > 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.
> >
>
> I've seen A7 positioned as ~25% slower than A9 at the same clocks, and a limited amount
> of benchmark data that supports this. Is there anything better about the memory system
> outside of the lower latency L2 cache? And where does it clock at 1.4GHz? ARM only says
> "> 1GHz." Exynos 5410 has a limit of 1.2GHz. Other SoCs I know of are 1GHz.
I guess it benefits mostly from the integrated low-latency L2. But there are likely other improvements throughout the memory system - an improved memory controller or prefetcher might help even more on an in-order than on A9R4.
In terms of frequency, it does 1.5GHz.
> > According to Samsungs ISSC graph, a quad A7 at 1.2GHz uses ~500mW, or just ~125mW per core.
> >
>
> The numbers I gave are also from Samsung (in their kernel source) so the two
> are in disagreement. It could be that those numbers are for a worse bin.
What do those numbers mean and how are they used?
> > 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.
> >
>
> Aside from that this would only apply to code that scales perfectly with core count it also sounds like
> it's vastly undershooting the benefit of HT. It's common for it to improve performance on Atom by 40, even
> 50%. That'd bring the needed clocks for Atom as low as 1.4GHz, which grossly changes the comparison.
Yes HT can give a good benefit, but it doesn't quite scale like a real core, so it's unlikely you always get 40-50%. We don't know the exact frequencies, but given that A7 is pretty fast (it may be closer to 1GHz A7 = 1GHz Atom) and might scale better than A9 due to its memory system, it is quite feasible for a quad A7 to beat even the fastest Atom.
Wilco
> Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on June 3, 2013 10:52 am wrote:
> > 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.
>
> No, but it's better than the Silvermont hardware ARM has access to, which is nothing.
True. But Intel's claims would not be correct if they are based just on the 5250 and assume no further improvement (tuning, big.Little, 20nm process change etc).
> > 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.
> >
>
> I've seen A7 positioned as ~25% slower than A9 at the same clocks, and a limited amount
> of benchmark data that supports this. Is there anything better about the memory system
> outside of the lower latency L2 cache? And where does it clock at 1.4GHz? ARM only says
> "> 1GHz." Exynos 5410 has a limit of 1.2GHz. Other SoCs I know of are 1GHz.
I guess it benefits mostly from the integrated low-latency L2. But there are likely other improvements throughout the memory system - an improved memory controller or prefetcher might help even more on an in-order than on A9R4.
In terms of frequency, it does 1.5GHz.
> > According to Samsungs ISSC graph, a quad A7 at 1.2GHz uses ~500mW, or just ~125mW per core.
> >
>
> The numbers I gave are also from Samsung (in their kernel source) so the two
> are in disagreement. It could be that those numbers are for a worse bin.
What do those numbers mean and how are they used?
> > 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.
> >
>
> Aside from that this would only apply to code that scales perfectly with core count it also sounds like
> it's vastly undershooting the benefit of HT. It's common for it to improve performance on Atom by 40, even
> 50%. That'd bring the needed clocks for Atom as low as 1.4GHz, which grossly changes the comparison.
Yes HT can give a good benefit, but it doesn't quite scale like a real core, so it's unlikely you always get 40-50%. We don't know the exact frequencies, but given that A7 is pretty fast (it may be closer to 1GHz A7 = 1GHz Atom) and might scale better than A9 due to its memory system, it is quite feasible for a quad A7 to beat even the fastest Atom.
Wilco
