[petsc-users] Does float128 helps ksp solver convergence

[Barry Smith]

On Sep 18, 2011, at 11:09 AM, Gong Ding wrote:

> Thanks, the synopsys TCAD product announced they use 128 bit float for numerically stable reason.
> I will try some more robust methods first.
> float128 seems not compatible with most external direct solvers.

Semiconductor  simulations are notoriously ill-conditioned; personally I would always run them in quad precision. But note that the extra precision does not make the iterative solvers any faster they will still be very slow without very good preconditioners (which are also notoriously difficult for semiconductors.)

I recommend contacting the SuperLU_Dist and MUMPS people and asking them to implement a quad precision version; I suspect it might not be difficult for SuperLU_Dist but since MUMPS lives on top of the rickety infrastructure of ScaLAPACK it is probably very difficult for them. So contact Sherrie Li or see about modifiying SuperLU_Dist yourself.

   Barry



>  
> On Sun, Sep 18, 2011 at 17:07, Gong Ding <gdiso at ustc.edu> wrote:
> Do you mean GMRES-like method will be benefited from float 128?
> 
> It could be, especially if you use a finite difference Jacobian. If you suspect precision issues, you can try more robust orthogonalization -ksp_gmres_modifiedgramschmidt or -ksp_type gcr, or you can just try using float128.
> 
>