By: jp (asdfasdf.delete@this.gmail.com), August 13, 2012 11:17 am
Room: Moderated Discussions
David Kanter (dkanter.delete@this.realworldtech.com) on August 13, 2012 9:50 am wrote:
> > > Speaking of
> > > power, SP vs. DP which you think consumes
>
> > more power?
> > > Could Kepler turbo on
> > > DP?
> > >
>
> > > K10 is
> > 745MHz/112.5W, x2880/1536=211W,
> x1.5T/(2*960*745M)=221W
> > > I guess
> > >
> > that big-K at
> 4.5T/1.5T is plausible at 1.8T/300W.
> >
> > Interesting notion
> >
> considering that they now seem to be doing the DP operations on separate units.
>
> > Perhaps they would able to do these operations at a higher clock?
>
> They
> are using the same datapath for GK110. IIRC, each SM has 3 sets of 16 FPUs.
> Two of those sets can be ganged together to execute 16 DP instructions.
>
> I'm
> not sure about GK104 though.
The initial hardware schematics I read through showed that the GK104 had a separate set of FP64 units.
But it makes much more sense that they would use the same datapath as they did with Fermi ( 192*(2/3) * ( (1 64-bit op)/(2 FPUs ) ) => 64 ).
>
> > When
> > measuring the actual power
> consumption of the GT240 we found that purely compute
> > bound applications
> didn't manage to max out the power consumption (50-60 % of
> > max power
> while reaching near peak GFLOP/s numbers). It was rather the bandwidth
> >
> bound applications that seemed to put more strain on the memory controllers
> that
> > were consuming the most power.
>
> The memory controller consumes a
> ton of power and is definitely required to max the chip out.
>
> DK
Yes, and the texture units seem to be really power hungry.
> > > Speaking of
> > > power, SP vs. DP which you think consumes
>
> > more power?
> > > Could Kepler turbo on
> > > DP?
> > >
>
> > > K10 is
> > 745MHz/112.5W, x2880/1536=211W,
> x1.5T/(2*960*745M)=221W
> > > I guess
> > >
> > that big-K at
> 4.5T/1.5T is plausible at 1.8T/300W.
> >
> > Interesting notion
> >
> considering that they now seem to be doing the DP operations on separate units.
>
> > Perhaps they would able to do these operations at a higher clock?
>
> They
> are using the same datapath for GK110. IIRC, each SM has 3 sets of 16 FPUs.
> Two of those sets can be ganged together to execute 16 DP instructions.
>
> I'm
> not sure about GK104 though.
The initial hardware schematics I read through showed that the GK104 had a separate set of FP64 units.
But it makes much more sense that they would use the same datapath as they did with Fermi ( 192*(2/3) * ( (1 64-bit op)/(2 FPUs ) ) => 64 ).
>
> > When
> > measuring the actual power
> consumption of the GT240 we found that purely compute
> > bound applications
> didn't manage to max out the power consumption (50-60 % of
> > max power
> while reaching near peak GFLOP/s numbers). It was rather the bandwidth
> >
> bound applications that seemed to put more strain on the memory controllers
> that
> > were consuming the most power.
>
> The memory controller consumes a
> ton of power and is definitely required to max the chip out.
>
> DK
Yes, and the texture units seem to be really power hungry.
