By: Peter Boyle (paboyle.delete@this.ph.ed.ac.uk), July 27, 2012 7:57 am
Room: Moderated Discussions
Moritz (better.delete@this.not.tell) on July 26, 2012 1:34 am wrote:
> The three entries BG/Q, Fermi, Cypress are nearly on a line.
> This suggests that
> one could extract the trade-off-factor of power/area.
> Under the circumstances
> common to those designs it seems to be:
> 1.35(mm)²/W
> Because fermi is in the
> middle this factor might be close to that of that design. Really Fermi is above
> and to the right of the curve suggested by BG/Q and Cypress which makes it a
> good compromise.
> For speculations this might do.
Nonsense.
The reason BG/Q has bigger area/flop is because it has a huge
cache.
BG/Q would have higher power efficiency AND higher area efficiency
if it dropped the huge cache (and MPI network) -- the BG/Q point would
move far to the right and slightly up if the cache were dropped to
the levels of their competitors.
However, it wouldn't be as capable for scientific computing -- and that
is the engineering trade-off.
Peter Boyle
> The three entries BG/Q, Fermi, Cypress are nearly on a line.
> This suggests that
> one could extract the trade-off-factor of power/area.
> Under the circumstances
> common to those designs it seems to be:
> 1.35(mm)²/W
> Because fermi is in the
> middle this factor might be close to that of that design. Really Fermi is above
> and to the right of the curve suggested by BG/Q and Cypress which makes it a
> good compromise.
> For speculations this might do.
Nonsense.
The reason BG/Q has bigger area/flop is because it has a huge
cache.
BG/Q would have higher power efficiency AND higher area efficiency
if it dropped the huge cache (and MPI network) -- the BG/Q point would
move far to the right and slightly up if the cache were dropped to
the levels of their competitors.
However, it wouldn't be as capable for scientific computing -- and that
is the engineering trade-off.
Peter Boyle
