Using single-precision for faster calculation

[septc]

Today I came across these articles about the use of single and mixed precisions (for molecular dynamics and electronic structure calculations):

• Routine Microsecond Molecular Dynamics Simulations with AMBER on GPUs
  https://pubs.acs.org/doi/10.1021/ct200909j

• Double Precision Is Not Needed for Many-Body Calculations: Emergent Conventional Wisdom
  https://pubs.acs.org/doi/10.1021/acs.jctc.8b00321

• Accelerating seminumerical Fock-exchange calculations using mixed single- and double-precision arithmethic
  https://aip.scitation.org/doi/full/10.1063/5.0045084

This article seems to be using “half-precision”:

• Mixed Precision Fermi-Operator Expansion on Tensor Cores from a Machine Learning Perspective
  https://pubs.acs.org/doi/10.1021/acs.jctc.1c00057

---

@RonShepard

Finally, as a practical matter, you don’t want to go through your codes and change back and forth between D and E exponents, or between “sp” and “dp” kind parameters. You want to write your code with something neutral, like “wp”, and then change the whole shebang by just defining what the “wp” kind value is in one place. This is one of the big advantages of modern fortran over other languages, so don’t just ignore that capability.

Up to now I used only double precision in my codes, so various parts uses d0 for floating-point literals… So first of all, I need to fix these literals to _rp for experiment (precisely as you mentioned above )

@rwmsu

To me the issue is not can you use single instead of double on a CPU but on current GPUs. There are $600 desktop graphics cards that can do 10 + TFLOPS in FP32 and 800 + Gflops in FP64 so if you have a code that can map efficiently to a GPU going with single precision can have a hugh payoff

Yes, GPU is very attractive and I think the motivation for the above-linked papers is probably GPU. Though I cannot use GPUs right now, I would like to try them once the program and calculations become more “stable” (possibly with OpenACC?)