Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on September 25, 2013 3:14 am wrote:
> none (none.delete@this.none.com) on September 25, 2013 2:50 am wrote:
> > anon (anon.delete@this.anon.com) on September 25, 2013 1:56 am wrote:
> > > Doug S (foo.delete@this.bar.bar) on September 24, 2013 8:44 pm wrote:
> > > > Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on September 24, 2013 6:17 pm wrote:
> > > > > Rubbish. In Anand's Calxeda test a 2.3GHz E5-2560L beat a quad 1.4GHz A9 by a factor of 4.8x
> > > > > single threaded. That means a 4GHz i7 would beat an 1.8GHz Calxeda node by ~6.5x. Given that
> > > > > A15 is significantly faster than A9, the i7 would beat A15 by about 3.5x. That means a 4GHz
> > > > > i7 would only be ~2.3x faster than an A7 at 1.8GHz. So they are pretty close already.
> > > >
> > > >
> > > > That's an awful lot of commutative math there, it is easy to trip up (or
> > > > deliberately obfuscate) with all that, especially without citations.
> > > >
> > > > But basically we know it is a steaming load of you know what, given the final claim that
> > > > the 2 watt SoC in a phone is only 2.3x slower than an actively cooled top bin Haswell.
> > > > Surely you can't be deluded enough to really believe that? If you do, you're basically
> > > > telling people they shouldn't bother to read anything you post ever again.
> > >
> > > If you compare some geekbench scores (integer, single threaded)
> > >
> > > http://browser.primatelabs.com/geekbench3/78873
> > > http://browser.primatelabs.com/geekbench3/79461
> > >
> > > Ignore the crypto, because I believe that's been highly optimized on x86 and
> > > uses crypto instructions. Take a look at some of the other sub tests.
> > >
> > > Geo mean of the non-crypto single thread int subtests gives
> > > Haswell a 4.8x advantage. Given the power consumption,
> > > caches, memory subsystems, and turbo frequencies, I think
> > > this is a really good performance by the A15, but it
> > > does only achieve 1/2 the IPC of the Haswell, which is a long way from Wilco's claim of IPC parity.
> >
> > Wasn't Wilco claiming Haswell vs Cyclone parity rather than Haswell vs A15?
>
> Correct.
>
> > http://browser.primatelabs.com/geekbench3/compare/81817?baseline=79461
>
> That looks like an i7 overclocked to 4.8GHz. Pretty much all i7-4770 results are 32-bit
> and in the 3500-3600 range. After a bit of browsing I found a 64-bit i7 result:


Hmm, I thought geekbench tried to compute the clock, but maybe that's wrong. Anyway, you maybe right, which does make the A15 a little faster relatively, but probably not down to 3.5x.

>
> http://browser.primatelabs.com/geekbench3/compare/81817?baseline=43547
>
> That suggests a 3.9GHz i7 is 2.8x faster than A7 on single threaded integer with a 3x clock advantage.

I still find the crypto tests questionable. SHA1 for example.