You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@mxnet.apache.org by Chris Olivier <cj...@gmail.com> on 2018/04/02 22:22:43 UTC
Fermi Architecture Support (pre-Kepler)
Cuda9 nvcc no longer supports Fermi (pre-Kepler) architectures (2.0, 2.1)
It appears that neither TensorFlow or PyTorch support Fermi.
Are there any objections (with justification) to continue support for Fermi
(20, 21)?
Re: Fermi Architecture Support (pre-Kepler)
Posted by Chris Olivier <cj...@gmail.com>.
Tomorrow I am going to nuke pre-Kepler support. Speak now or forever hold
your peace.
On Tue, Apr 3, 2018 at 10:08 AM Anirudh <an...@gmail.com> wrote:
> +1 cudnn doesn't support this either.
>
> On Tue, Apr 3, 2018, 2:03 AM kellen sunderland <
> kellen.sunderland@gmail.com>
> wrote:
>
> > I doubt practically speaking that many pre-Keplar devices would be that
> > useful for MXNet anyway. +1 to dropping them and making our codebase
> more
> > readable / maintainable.
> >
> > On Tue, Apr 3, 2018 at 12:23 AM, Marco de Abreu <
> > marco.g.abreu@googlemail.com> wrote:
> >
> > > Definitely in favour of dropping support!
> > >
> > > On Tue, Apr 3, 2018 at 12:22 AM, Chris Olivier <cj...@gmail.com>
> > > wrote:
> > >
> > > > Cuda9 nvcc no longer supports Fermi (pre-Kepler) architectures (2.0,
> > 2.1)
> > > > It appears that neither TensorFlow or PyTorch support Fermi.
> > > >
> > > > Are there any objections (with justification) to continue support for
> > > Fermi
> > > > (20, 21)?
> > > >
> > >
> >
>
Re: Fermi Architecture Support (pre-Kepler)
Posted by Anirudh <an...@gmail.com>.
+1 cudnn doesn't support this either.
On Tue, Apr 3, 2018, 2:03 AM kellen sunderland <ke...@gmail.com>
wrote:
> I doubt practically speaking that many pre-Keplar devices would be that
> useful for MXNet anyway. +1 to dropping them and making our codebase more
> readable / maintainable.
>
> On Tue, Apr 3, 2018 at 12:23 AM, Marco de Abreu <
> marco.g.abreu@googlemail.com> wrote:
>
> > Definitely in favour of dropping support!
> >
> > On Tue, Apr 3, 2018 at 12:22 AM, Chris Olivier <cj...@gmail.com>
> > wrote:
> >
> > > Cuda9 nvcc no longer supports Fermi (pre-Kepler) architectures (2.0,
> 2.1)
> > > It appears that neither TensorFlow or PyTorch support Fermi.
> > >
> > > Are there any objections (with justification) to continue support for
> > Fermi
> > > (20, 21)?
> > >
> >
>
Re: Fermi Architecture Support (pre-Kepler)
Posted by kellen sunderland <ke...@gmail.com>.
I doubt practically speaking that many pre-Keplar devices would be that
useful for MXNet anyway. +1 to dropping them and making our codebase more
readable / maintainable.
On Tue, Apr 3, 2018 at 12:23 AM, Marco de Abreu <
marco.g.abreu@googlemail.com> wrote:
> Definitely in favour of dropping support!
>
> On Tue, Apr 3, 2018 at 12:22 AM, Chris Olivier <cj...@gmail.com>
> wrote:
>
> > Cuda9 nvcc no longer supports Fermi (pre-Kepler) architectures (2.0, 2.1)
> > It appears that neither TensorFlow or PyTorch support Fermi.
> >
> > Are there any objections (with justification) to continue support for
> Fermi
> > (20, 21)?
> >
>
Re: Fermi Architecture Support (pre-Kepler)
Posted by Marco de Abreu <ma...@googlemail.com>.
Definitely in favour of dropping support!
On Tue, Apr 3, 2018 at 12:22 AM, Chris Olivier <cj...@gmail.com>
wrote:
> Cuda9 nvcc no longer supports Fermi (pre-Kepler) architectures (2.0, 2.1)
> It appears that neither TensorFlow or PyTorch support Fermi.
>
> Are there any objections (with justification) to continue support for Fermi
> (20, 21)?
>