By: David Kanter (dkanter.delete@this.realworldtech.com), May 18, 2013 7:50 am
Room: Moderated Discussions
> > That's not trickery, that's life. Intel has better process technology and is able to
> > hit higher clock speeds.
>
> It's trickery when you use a slow CPU on purpose when much faster CPUs are available.
> And in terms of frequency ARM has caught up dramatically in recent years. I expect ARM
> to pull ahead in frequency with Tegra 4i, 20nm A15's and the first 64-bit ARMs.
This is ridiculous. First, the 20nm process node at TSMC, GF and Samsung will be about the same performance as the 28nm node. Rephrasing: There is no reason to expect 20nm designs to be higher frequency than 28nm ones. I think this goes to prove my point quite eloquently, which is process matters.
Second, I'd like to emphasize that the DVFS used in most mobile devices is terribly primitive. Just look at the poor gains in frequency that you see in Tegra 3 when reducing the number of cores. But it's hard to tell what is a function of the CPU core design vs. the implementation; both can play a role.
Third, I don't think the A15 frequency is very good at low power, based on available data for current implementations. The power testing I've seen from Intel (which I trust, having seen the setup and instrumentation) shows that the 1.7GHz dual A15 from Samsung consumes about 1.5W/core when unconstrained by power management. Frankly, I think that 1W is about the limit for all cores when running in a smartphone...
To get from 1.5W/core down to 500mW/core will require a huge sacrifice in frequency. If we assume P = k*V2*F, that suggests that peak frequency would need to drop down to 1.1GHz.
1GHz is not an impressive frequency today, and is not 'catching up'. The Tegra 4i is a somewhat different story, as it's still based on the A9.
But the fundamental problem illustrated by the above example is that frequency isn't an abstract quantity. Frequency @ power is a meaningful metric and I'd like to see more clear disclosure from the ARM-based vendors in particular. E.g., yes Tegra 4i can run at 2.3GHz, but with how many cores and at what power consumption in the CPU block and at the platform level? How does it look with 2 cores? With all 4 cores? And how does power change as you sweep across different frequencies? How does the frequency change for different durations of workloads (e.g., how does it drop down over time)?
I think if you were to really look at the complete picture here and start asking those questions, you'd really see that Apple, Intel and Qualcomm get significant benefits from their custom CPU cores (and process technology for Intel).
David
> > hit higher clock speeds.
>
> It's trickery when you use a slow CPU on purpose when much faster CPUs are available.
> And in terms of frequency ARM has caught up dramatically in recent years. I expect ARM
> to pull ahead in frequency with Tegra 4i, 20nm A15's and the first 64-bit ARMs.
This is ridiculous. First, the 20nm process node at TSMC, GF and Samsung will be about the same performance as the 28nm node. Rephrasing: There is no reason to expect 20nm designs to be higher frequency than 28nm ones. I think this goes to prove my point quite eloquently, which is process matters.
Second, I'd like to emphasize that the DVFS used in most mobile devices is terribly primitive. Just look at the poor gains in frequency that you see in Tegra 3 when reducing the number of cores. But it's hard to tell what is a function of the CPU core design vs. the implementation; both can play a role.
Third, I don't think the A15 frequency is very good at low power, based on available data for current implementations. The power testing I've seen from Intel (which I trust, having seen the setup and instrumentation) shows that the 1.7GHz dual A15 from Samsung consumes about 1.5W/core when unconstrained by power management. Frankly, I think that 1W is about the limit for all cores when running in a smartphone...
To get from 1.5W/core down to 500mW/core will require a huge sacrifice in frequency. If we assume P = k*V2*F, that suggests that peak frequency would need to drop down to 1.1GHz.
1GHz is not an impressive frequency today, and is not 'catching up'. The Tegra 4i is a somewhat different story, as it's still based on the A9.
But the fundamental problem illustrated by the above example is that frequency isn't an abstract quantity. Frequency @ power is a meaningful metric and I'd like to see more clear disclosure from the ARM-based vendors in particular. E.g., yes Tegra 4i can run at 2.3GHz, but with how many cores and at what power consumption in the CPU block and at the platform level? How does it look with 2 cores? With all 4 cores? And how does power change as you sweep across different frequencies? How does the frequency change for different durations of workloads (e.g., how does it drop down over time)?
I think if you were to really look at the complete picture here and start asking those questions, you'd really see that Apple, Intel and Qualcomm get significant benefits from their custom CPU cores (and process technology for Intel).
David