GCN - TFLOP DP

EBFE (x.delete@this.y.com) on August 13, 2012 7:45 pm wrote:
> jp (asdasdf.delete@this.gmail.com) on August 13, 2012 2:02 am wrote:
> > EBFE
> (x.delete@this.y.com) on August 12, 2012 8:27 pm wrote:
> > > jp
> >
> (asdfasdf.delete@this.gmail.com) on August 10, 2012 1:23 am wrote:
> > >
> >
> > Eric
> > > (eric.kjellen.delete@this.gmail.com) on August 9,
> 2012 5:12 pm
> > wrote:
> > > >
> > > >
> > > > jp
> (asdfasdf.delete@this.gmail.com) on
> > August 9, 2012 1:58 pm
> > >
> wrote:
> > > > > >
> > > > EBFE
> >
> > > >
> > (x.delete@this.y.com) on July 25,
> > > 2012 3:25 am
> >
> wrote:
> > > > > >
> > > > > GCN
> > > > >
> >
> > >
> > > >
> > >
> > > > > >
> >
> > > > > Yes I
> > > >
> > > > didn't see
> >
> > > AMD:s
> > > GCN arch in the charts?
> > > >
> > >
> >
> > > > > >
> > > > > GCN is the
> > >
> first
> > >
> > > single chip ever to
> > > > > > push
> 1 TFLOP DP. And the new
> >
> > >
> > > > > professional
>
> > > > edition has
> > > > >
> > > ECC.
> > >
> > >
> > > >
> > > > > >
> > > > >
>
> >
> > > >
> > >
> >
> http://www.hpcwire.com/hpcwire/2012-08-08/amd_unveils_teraflop_gpu_with_
> >
> >
> > >
> > > >
> > > > >
> > > > >
>
> > ecc_support.html?featured=top
> > > > >
> > > >
>
> > > > That's
> > very impressive,
> > > > particularly
> with TDP
> > > > > at 274W.
> >
> > > Though we should
> probably expect Tesla K20
> > > > to arrive soon and
> > deliver
>
> > >
> > > > > higher performance in a comparable power
>
> > >
> > > envelope. I wonder
> > > why AMD hasn't
> >
> > > > released or
> > announced an HPC oriented
> > > >
> compute
> > > card based on the same GPU.
> > Could
> > > >
> > the reason be that AMD
> > > software
> > > > tools
> >
> are still immature in comparison with
> > > > > Nvidia's
> > >
> and
> > Intel's
> > > > offerings?
> > > >
> > > >
>
> > > > Well I've been
> > working with
> > > their tools using
> OpenCL on AMD
> > > > cards and it is
> > currently far more
> >
> > immature when compared to working with CUDA
> > >
> > > on
> Nvidia cards when it comes
> > > to driver reliability, developer
> >
> environment,
> > > > documentation, and feature
> > > set.
> >
> >
> > > Can
> > you elaborate it? esp. 'feature set'
> >
> > On
> feature set:
> >
> > - Direct access to
> > texture features for fast
> interpolations
> > - Warp voting functions
> > - certain
> > floating
> point atomic instructions*
> > - Register shuffle instructions (direct
> >
> exchange of data via registers instead of via smem )
> > - GPUDirect, DMA
> between
> > GPUs
> >
> > I'm sure I can lengthen this list if I spend
> some more time on it.
> >
> > *
> > ex atomicAdd(float* ), might have
> been introduced recently
> >
> Thanks. But what is this?
> > - Direct access
> to texture features for fast interpolations

You can perform interpolation directly in the texture hardware units. Instead of just fetching data from the texture cache I can also go ahead and directly do a linear interpolation at no extra cost and without occupying the shader units.