[QE-developers] Is an fp32 version of quantum espresso possible?

[Paolo Giannozzi]

On 3/23/26 17:43, Yash Rachamalla wrote:
 > > Greetings,
> 
> I am an undergraduate student trying to learn quantum espresso. I 
> installed it on my workstation but noticed my nvidia gpu was only 
> drawing 40w of power at 100% usage. I figured out it was because quantum 
> espresso only uses fp64 and not fp32. On enterprise hardware this isn't 
> a problem since fp64 is only half the throughput of fp32, but consumer 
> gpu are designed for graphics. On consumer nvidia gpus since kepler, 
> fp64 has 1/64th the throughput of fp32.
> 
> I was wondering if making an fp32 build is feasible? I am not planning 
> any publications, and I don't really need ultraprecise calculations. I 
> am just looking for a quick performance uplift.
> 
> Yash

Hi, there have been some attempts in the past to use single-precision 
FFTs. I am not sure the modified code is in a usable form right now. For 
sure there will be more such attempts in the future, either to switch 
some parts of the calculation to single precision, or to use double 
precision emulation, but it is unlikely that this will work on consumer GPUs

Paolo
-- 
Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche,
Univ. Udine, via delle Scienze 206, 33100 Udine Italy, +39-0432-558216