> It sounds like the "secret sauce" for this collaboration includes a set of numerical libraries.
Indeed. There are really only a limited set of (physics) choices when making these libraries. As long as the discretization you pick goes to QCD in the continuum limit, you can make whatever choices you want. Some choices lead to faster convergence, or easier numerics, or better symmetry, or whatever---at that point it's a cost/benefit analysis. But if your discretization ('lattice action') goes is in the QCD universality class ('has the right continuum limit') you're guaranteed to get the right answer as long as you can extrapolate to the continuum.
> It's a bit sad that there's so little glory in writing better numerical libraries.
Agreed, but physics departments (by and large) award tenure for doing physics, not for doing computer science. It's hard to get departments to say "yes, your expertise in optimizing GPU code is enough to get you on the tenure track".
> It's a common problem across the sciences. [...] I can believe they'd be reluctant to share.
The larger community does center around common codes. The biggest players are
Indeed. There are really only a limited set of (physics) choices when making these libraries. As long as the discretization you pick goes to QCD in the continuum limit, you can make whatever choices you want. Some choices lead to faster convergence, or easier numerics, or better symmetry, or whatever---at that point it's a cost/benefit analysis. But if your discretization ('lattice action') goes is in the QCD universality class ('has the right continuum limit') you're guaranteed to get the right answer as long as you can extrapolate to the continuum.
> It's a bit sad that there's so little glory in writing better numerical libraries.
Agreed, but physics departments (by and large) award tenure for doing physics, not for doing computer science. It's hard to get departments to say "yes, your expertise in optimizing GPU code is enough to get you on the tenure track".
> It's a common problem across the sciences. [...] I can believe they'd be reluctant to share.
The larger community does center around common codes. The biggest players are
USQCD http://usqcd-software.github.io/ quda http://lattice.github.io/quda/ grid https://github.com/paboyle/Grid/
but there are others, and there are private codes (like BMW's) too.
As part of the SciDAC program and now exascale initiative the DOE does fund a few software-focused national lab jobs. But not many.