By: Wilco (Wilco.Dijkstra.delete@this.ntlworld.com), July 11, 2013 6:40 am
Room: Moderated Discussions
none (none.delete@this.none.com) on July 11, 2013 5:58 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on July 11, 2013 5:33 am wrote:
> [...]
> > I don't follow. How disassembly tells us anything about denormals handling?
> > It's in control flags that we can't see.
>
> It doesn't. But as I had read that some people thought that the Android x86 version was
> using x87 (the Windows 32-bit version does use x87), I thought it made sense to say it.
>
> > > Now if you can also explain with denormals why CT+ is beaten by A15 and
> > > Krait even for single threaded integer tasks, I'm very interested :-)
> > >
> >
> > It's beaten by the margin that one would expect:
> > http://browser.primatelabs.com/geekbench2/compare/2122693/1970335
> > http://browser.primatelabs.com/geekbench2/compare/2147177/1970335
> > The difference is probably very similar to the difference in power consumption. Which, of course, means
> > that A15 and Krait are better CPUs since they got equal perf/Watt with higher absolute performance.
>
> Which makes me think I have always found the Lua score odd. ARM chips are doing very
> well on it (even a Cortex-A9 @ 1.4 GHz crushes a CT+ @ 2.0 GHz), and I wonder why.
>
> http://browser.primatelabs.com/geekbench2/compare/2085507/1970892
>
> The sum of the single-threaded benches is 4482 for the A9 and 4371 for the Z2580.
Yes even an old slow A9 beats the latest and fastest Atom on single threaded integer despite a significant memory system disadvantage - and no denormals to take the blame :-)
Note: use geomean to compute the average, not a sum - another thing AnTuTu gets wrong and a reason for allowing their scores to be easily gamed by a single outlier.
Wilco
> Michael S (already5chosen.delete@this.yahoo.com) on July 11, 2013 5:33 am wrote:
> [...]
> > I don't follow. How disassembly tells us anything about denormals handling?
> > It's in control flags that we can't see.
>
> It doesn't. But as I had read that some people thought that the Android x86 version was
> using x87 (the Windows 32-bit version does use x87), I thought it made sense to say it.
>
> > > Now if you can also explain with denormals why CT+ is beaten by A15 and
> > > Krait even for single threaded integer tasks, I'm very interested :-)
> > >
> >
> > It's beaten by the margin that one would expect:
> > http://browser.primatelabs.com/geekbench2/compare/2122693/1970335
> > http://browser.primatelabs.com/geekbench2/compare/2147177/1970335
> > The difference is probably very similar to the difference in power consumption. Which, of course, means
> > that A15 and Krait are better CPUs since they got equal perf/Watt with higher absolute performance.
>
> Which makes me think I have always found the Lua score odd. ARM chips are doing very
> well on it (even a Cortex-A9 @ 1.4 GHz crushes a CT+ @ 2.0 GHz), and I wonder why.
>
> http://browser.primatelabs.com/geekbench2/compare/2085507/1970892
>
> The sum of the single-threaded benches is 4482 for the A9 and 4371 for the Z2580.
Yes even an old slow A9 beats the latest and fastest Atom on single threaded integer despite a significant memory system disadvantage - and no denormals to take the blame :-)
Note: use geomean to compute the average, not a sum - another thing AnTuTu gets wrong and a reason for allowing their scores to be easily gamed by a single outlier.
Wilco