Maynard Handley (name99.delete@this.name99.org) on September 28, 2013 9:00 am wrote:
> anon (anon.delete@this.anon.com) on September 28, 2013 3:22 am wrote:
> > Maynard Handley (name99.delete@this.name99.org) on September 27, 2013 5:03 pm wrote:
> > > someone (someone.delete@this.somewhere.com) on September 27, 2013 12:00 pm wrote:
> > > > Maynard Handley (name99.delete@this.name99.org) on September 27, 2013 11:30 am wrote:
> > > >
> > > > > > The situation today is very much different from that late 80's and '90s'. Simply
> > > > > > saying ARM will win because it is coming up from below just like x86 did to the
> > > > > > server RISCs is grossly simplistic as well as almost certainly wrong. The money
> > > > > > situation is very different and that determines the winner.
> > > > >
> > > > > While the money involved is relevant, the more relevant issue is the engineering involved.
> > > > > The flaw with x86 is the insane complexity of the designs. This complexity means
> > > > > - seven years to design a CPU (and so high latency in responding to changing markets)
> > > > > - multiple design teams are required (and have to be paid) to create a new design every year
> > > > > ARM is so much simpler (and will be even simpler when people start making ARMv8 only chips) that one can
> > > > > mount a serious challenge to Intel from below at VASTLY less expense, and on a VASTLY shorter timescale.
> > > >
> > > > Decades ago (MIPS R2000 vs 386 say) RISC had a distinct advantage in terms
> > > > of results for design effort. For any market that needs more than a low frequency
> > > > in-order scalar processor the difference drops down to noise levels today as
> > > > performance requirements rise to even just current smart phone levels.
> > > >
> > > > Saying an A7 core was vastly easier for Apple to design than the Bay Trail core
> > > > was for Intel to design is just sheer nonsense. High performance x86 is a long
> > > > solved problem (at Intel if nowhere else) whose solutions continuously improve
> > > > based on accumulated knowledge. The portion of a Baytrail SoC dedicated to
> > > > x86 decoding that the A7 SoC lacks is ridiculously infinitesimal in terms of the
> > > > over all device with multiprocessor cache coherency, OOOE execution cores,
> > > > dynamic power management, and all of the non-CPU integrated functionality.
> > >
> > > So Apple can create a design that is competitive with Intel, with what appears
> > > to be a substantially smaller team, certainly with vastly less experience, and
> > > far faster than Intel --- but there is no meaningful difference in complexity?
> >
> > What is the size of the Silvermont design team versus A7, and how long did they take to design?
> >
>
> I don't know the size of the teams (and I expect Apple will never release this info) but we do know the rate
> of Apple's progress, and we also know the data point I have stated many times --- a desktop Intel CPU takes about
> 7 years to get from concept to shipping. We saw this confirmed pretty much in Atom, where (apparently without
> the benefit of parallel design teams) it took them about six years to get from one micro-architecture to the
> next, with the intervening years spent on annual spins that improved the packaging or shrank the die.
> We have AMD's constant failure in this area to meet Intel's performance levels as a second data point.
>
> Look, if you want to insist on the (IMHO insane) claim that it is every bit as easy to design
> a high performance x86 device as to design a high-performance AArch64 device, be my guest; I
> don't give damn. Tech seems to bring out crazy delusions in people, you are welcome to yours.

It's generally accepted if you design any bespoke architecture in any ISA it takes about 4 years and if you license cores about half that. Silvermont contains 2008 Westmere technology and Apple bought PA-Semi in 2008. Apple A6 debuted in 2012 and Silvermont only a year later held up by waiting for 32nm Atom SoC to tip up, 22nm LP to become available and the forced software change from Meego to Android. Intel is now on a tick-tock with Atom so 2014 14nm Airmont will be followed by 2015 14nm Goldmont and 2016 10nm Goldmont.

http://en.wikipedia.org/wiki/P.A._Semi