By: Emil Briggs (me.delete@this.nowherespam.com), October 20, 2012 8:34 am
Room: Moderated Discussions
Emil Briggs (me.delete@this.nowherespam.com) on October 20, 2012 8:33 am wrote:
> > >
> > > Just for reference. I'm not
> > anonymous and I could
> care less
> > > about tenure.
> > >
> > That's a good
> >
> start.
> >
> > > A lot of supercomputer time gets used for density
> functional
> >
> > > calculations. And it has produced valid and useful
> results that are
> > interesting
> > > both commercially and from a
> pure science perspective. Indeed
> > such calculations
> > > have helped
> produce that desktop that you talk about
> > since a detailed
> > >
> understanding of the physical mechanisms involved is
> > important to
> improving
> > > semiconductor processes.
> > >
> > I was about two
>
> > say that, well, you're the first person who hasn't responded to the
> effect that
> > we are using horrible approximations with these big boxes to
> get answers with
> > unknown and unknowable accuracy. But then I remembered
> two things. I have
> > previously been informed by what I take to be a solid
> state physicist at IBM
> > Watson that they were calculating solid state
> problems with answers accurate to
> > five digits. Then I also remembered
> that one of the big open problems of big
> > box computing was that
> predictions of electron mobility were off by a factor of
> > five. Maybe
> that problem has been fixed, or maybe it's only necessary to get
> > the sign
> of an effect correct (if you strain the lattice, the electron mobility
> >
> improves--only, as Intel discovered, to its dismay, not as much as they had
>
> > hoped). That you regard these gigantic calculations as a necessary tool
> of your
> > trade and that you are passionate about it is clear. That I have
> exhausted the
> > entire extent of my detailed knowledge of the subject is
> also clear. That's not
> > to say that my understanding of the problem
> couldn't be improved, only, not in
> > the context of a discussion like the
> current one. There's also a lot of
> > gigantic cluster time being used to
> do computer animation and to do inverse
> > scattering for oil exploration,
> and, at one time, someone in a position to know
> > said that he knew of
> commercial clusters that had no interest in being on the
> > Top 500 list but
> that would have blown it away. The difference is that those
> > clusters are
> being funded with private risk capital, albeit with tax subsidies
> >
> probably in both cases, and the decision to build them was clearly not driven
> by
> > egos and academic/national lab politics.
> >
>
> There are physically
> measurable quantities in some solid state problems that can be calculated to a
> high degree of precision. But 5 digits is not necessary for the results to be
> useful and 2 or 3 digits is possible for a wide range of problems. DFT has even
> been successful in predicting some physical phenomena before they were
> experimentally observed which is a pretty good test of the validity of a
> particular methodology.
>
> > > And there are DFT methods that
> > scale
> well
> > > computationally on supercomputers. A bigger limitation is
> that
> > most DFT methods
> > > exhibit O(N^3) or worse scaling with the
> number of atoms
> > being simulated which
> > > limits things to a few
> thousand atoms on the
> > current generation of
> > > supercomputers.
> To study larger systems (and there
> > are plenty of commercial and
> >
> > scientific reasons to want to do so) will
> > require new
> algorithms.
> > >
> > Well, now there are two of you and a paper who
>
> > appear to believe in perpetual motion.
> >
>
> I not sure what point
> you're trying to make here. DFT is based on first principles and has a very good
> track record but it's computationally very expensive. You can do a lot of useful
> work with it on systems that range up to a few thousand atoms but with O(N^3)
> scaling you can easily use up any feasible increase in computer power without
> extending the size range you can model very far. Hence the efforts to improve
> the algorithmic scaling. It's not a solved problem yet but it's not perpetual
> motion.
>
> > >
> > > Just for reference. I'm not
> > anonymous and I could
> care less
> > > about tenure.
> > >
> > That's a good
> >
> start.
> >
> > > A lot of supercomputer time gets used for density
> functional
> >
> > > calculations. And it has produced valid and useful
> results that are
> > interesting
> > > both commercially and from a
> pure science perspective. Indeed
> > such calculations
> > > have helped
> produce that desktop that you talk about
> > since a detailed
> > >
> understanding of the physical mechanisms involved is
> > important to
> improving
> > > semiconductor processes.
> > >
> > I was about two
>
> > say that, well, you're the first person who hasn't responded to the
> effect that
> > we are using horrible approximations with these big boxes to
> get answers with
> > unknown and unknowable accuracy. But then I remembered
> two things. I have
> > previously been informed by what I take to be a solid
> state physicist at IBM
> > Watson that they were calculating solid state
> problems with answers accurate to
> > five digits. Then I also remembered
> that one of the big open problems of big
> > box computing was that
> predictions of electron mobility were off by a factor of
> > five. Maybe
> that problem has been fixed, or maybe it's only necessary to get
> > the sign
> of an effect correct (if you strain the lattice, the electron mobility
> >
> improves--only, as Intel discovered, to its dismay, not as much as they had
>
> > hoped). That you regard these gigantic calculations as a necessary tool
> of your
> > trade and that you are passionate about it is clear. That I have
> exhausted the
> > entire extent of my detailed knowledge of the subject is
> also clear. That's not
> > to say that my understanding of the problem
> couldn't be improved, only, not in
> > the context of a discussion like the
> current one. There's also a lot of
> > gigantic cluster time being used to
> do computer animation and to do inverse
> > scattering for oil exploration,
> and, at one time, someone in a position to know
> > said that he knew of
> commercial clusters that had no interest in being on the
> > Top 500 list but
> that would have blown it away. The difference is that those
> > clusters are
> being funded with private risk capital, albeit with tax subsidies
> >
> probably in both cases, and the decision to build them was clearly not driven
> by
> > egos and academic/national lab politics.
> >
>
> There are physically
> measurable quantities in some solid state problems that can be calculated to a
> high degree of precision. But 5 digits is not necessary for the results to be
> useful and 2 or 3 digits is possible for a wide range of problems. DFT has even
> been successful in predicting some physical phenomena before they were
> experimentally observed which is a pretty good test of the validity of a
> particular methodology.
>
> > > And there are DFT methods that
> > scale
> well
> > > computationally on supercomputers. A bigger limitation is
> that
> > most DFT methods
> > > exhibit O(N^3) or worse scaling with the
> number of atoms
> > being simulated which
> > > limits things to a few
> thousand atoms on the
> > current generation of
> > > supercomputers.
> To study larger systems (and there
> > are plenty of commercial and
> >
> > scientific reasons to want to do so) will
> > require new
> algorithms.
> > >
> > Well, now there are two of you and a paper who
>
> > appear to believe in perpetual motion.
> >
>
> I not sure what point
> you're trying to make here. DFT is based on first principles and has a very good
> track record but it's computationally very expensive. You can do a lot of useful
> work with it on systems that range up to a few thousand atoms but with O(N^3)
> scaling you can easily use up any feasible increase in computer power without
> extending the size range you can model very far. Hence the efforts to improve
> the algorithmic scaling. It's not a solved problem yet but it's not perpetual
> motion.
>
