By: anon (anon.delete@this.anon.com), April 21, 2015 11:24 pm
Room: Moderated Discussions
Exophase (exophase.delete@this.gmail.com) on April 21, 2015 9:25 pm wrote:
> Maynard Handley (name99.delete@this.name99.org) on April 21, 2015 8:57 pm wrote:
> > You're discussing three (apparently) orthogonal ideas: QWFETs, III-V (or Germanium) transistors,
> > and 10nm. Are there good reasons to believe these HAVE to be yoked together? Is it possible,
> > for example (especially given Intel's difficulties and delays at 14nm) that they produce Skylake
> > on 14nm, then Cannonlake on 14nm with QWFET and III-V? This would be something of a departure
> > from the standard tick-tock model, but that model is merely a guideline.
> > Doing this would allow them to maintain the microarchitecture cadence rather than being forced
> > to slow down to match process upgrades, would allow them to debug QWFET issues on a somewhat better
> > known process, and would give them another 18 months or so to deal with 10nm issues.
> >
>
> A big motivator for Moore's Law has been lowering cost per transistor. Outside of being an easier
> design target for straightening out the new process, the financial allure of the tick is allowing
> them to sell (what's close to) the previous design at a lower cost. Lower power consumption or higher
> performance are fringe benefits in comparison.
This is no longer exactly true for high performance CPUs. If you halved the power consumption of a device, then you could design something equivalent to previous design in performance using fewer transistors.
It's especially not true for mobile when you look at the whole device cost, you also reduce cost of battery and thermal design.
> At least for the big cores where Intel has a lock
> the markets they're playing in (even the tablets are something of a class unto themselves)
Transistors are so cheap now that 4-8 multithreaded cores and 8-20MB of cache is common on a desktop CPU, and laptops are not so far behind.
Selling close to previous design for lower cost has never been Intel's game. They've always pushed to improve on designs and maintain prices.
>
> Maybe such a thing would be appealing for Atom but I'm skeptical it'd be something they'd do
> for Cannonlake, something that'd make it (possibly significantly) more expensive than Skylake.
> Maynard Handley (name99.delete@this.name99.org) on April 21, 2015 8:57 pm wrote:
> > You're discussing three (apparently) orthogonal ideas: QWFETs, III-V (or Germanium) transistors,
> > and 10nm. Are there good reasons to believe these HAVE to be yoked together? Is it possible,
> > for example (especially given Intel's difficulties and delays at 14nm) that they produce Skylake
> > on 14nm, then Cannonlake on 14nm with QWFET and III-V? This would be something of a departure
> > from the standard tick-tock model, but that model is merely a guideline.
> > Doing this would allow them to maintain the microarchitecture cadence rather than being forced
> > to slow down to match process upgrades, would allow them to debug QWFET issues on a somewhat better
> > known process, and would give them another 18 months or so to deal with 10nm issues.
> >
>
> A big motivator for Moore's Law has been lowering cost per transistor. Outside of being an easier
> design target for straightening out the new process, the financial allure of the tick is allowing
> them to sell (what's close to) the previous design at a lower cost. Lower power consumption or higher
> performance are fringe benefits in comparison.
This is no longer exactly true for high performance CPUs. If you halved the power consumption of a device, then you could design something equivalent to previous design in performance using fewer transistors.
It's especially not true for mobile when you look at the whole device cost, you also reduce cost of battery and thermal design.
> At least for the big cores where Intel has a lock
> the markets they're playing in (even the tablets are something of a class unto themselves)
Transistors are so cheap now that 4-8 multithreaded cores and 8-20MB of cache is common on a desktop CPU, and laptops are not so far behind.
Selling close to previous design for lower cost has never been Intel's game. They've always pushed to improve on designs and maintain prices.
>
> Maybe such a thing would be appealing for Atom but I'm skeptical it'd be something they'd do
> for Cannonlake, something that'd make it (possibly significantly) more expensive than Skylake.
