anon (anon.delete@this.anon.com) on November 26, 2013 6:42 pm wrote:
> none (none.delete@this.none.com) on November 26, 2013 10:18 am wrote:
> > Michael S (already5chosen.delete@this.yahoo.com) on November 26, 2013 9:06 am wrote:
> > > none (none.delete@this.none.com) on November 26, 2013 7:20 am wrote:
> > > > anon (anon.delete@this.anon.com) on November 26, 2013 6:53 am wrote:
> > > > > Eric (eric.kjellen.delete@this.gmail.com) on November 26, 2013 4:05 am wrote:
> > > > > > Linus Torvalds (torvalds.delete@this.linux-foundation.org) on November 25, 2013 12:29 pm wrote:
> > > > > > > And the old Atom really was pretty bad at some general-purpose stuff. That VR-zone link
> > > > > > > says KNL is 72 modified Silvermont cores, so it should be much better in that regard.
> > > > > > >
> > > > > > > Linus
> > > > > >
> > > > > > What I am mostly interested in is exactly how they are modified
> > > > > > with regard to vector processing. The Knights
> > > > > > Corner cores and those of its predecessors were basically P54C derivatives with an added 512-bit vector
> > > > > > unit (for 16 DP FMA flops per core and cycle) but as we
> > > > > > speculated in a different branch of the thread (what
> > > > > > is now "3-TFLOPS-DGEMM") 72 Silvermont cores with a single 512-bit vector unit will not be enough to give
> > > > > > the cited 3 DP peak TFLOPS at a reasonable clock frequency (they would have to run at 2.6 GHz).
> > > > >
> > > > > Why is 2.6GHz an unreasonable frequency?
> > > > >
> > > > > Intel's 14nm silvermont derivative is reported to have 2.7GHz clock.
> > > >
> > > > Try to put 72 of them along with 2 wide vector units and you'll see that doesn't seem
> > > > reasonable even on 14nm :-)
> > >
> > > Yes, judged by Avaton C2750, when running at max. frequency 72
> > > Silvermont cores alone, without VPUs, would eat over 150W.
>
> How do you calculate that? Compare between C2550 and C2750, and you add 6W TDP for 4
> cores (plus 2MB L3, inteconnect, etc, but ignore that). Then reduce power by 30% for
> 14nm. That comes out at 1.05W per core at 2.4GHz (presumably). Increase frequency to
> 2.6 probably brings power up significantly, but it seems 72 cores could fit into 100W.

You think the VPU won't come with increased power? Do you know that on Haswell as soon
as the AVX unit gets used voltage is increased (at least that's what happen on my
system)? Have you ever seen Haswell power consumption skyrocket as soon as AVX is
heavily used? And that's with a 256-bit SIMD unit.

Then you have to add eDRAM, interconnect, etc.

> > > But why did you mention "2 wide vector units"? At 2.7 GHz one unit per core will do the job.
> >
> > You're right. But current rumors point to Silvermont + dual VPU, so in that context
> > saying 14nm next gen Atom can reach 2.7 GHz kind of assumes these rumors are correct,
> > so 2 VPUs :-)
>
> No it doesn't! The whole point is to derive how they can achieve
> 3 TFLOPS. If it runs at 2.6GHz, then it requires 1 vector unit.

OK, if you insist, but IMHO that still won't fit.