CCL:G: [CCL]Re: CCL:G: Gaussian 09 and 16 compared for mundane calculations

[Igors Mihailovs <igorsm]=[cfi.lu.lv>]

     Hello Dr. DeYoung,
       
       Personally I have an overall impression that Gaussian 16 is a bit
       faster either because of algorithm improvements or maybe because
       it utilizes AVX2 CPU instructions available on our newer nodes.
       Although I have now checked and for small molecules Gaussian 09 at
       HF level somehow is consistently a bit faster for the second
       hyperpolarizability.
       
       In regards to difference in results. The main difference in
       defaults is the integration grid for DFT which is now UltraFine
       and not Fine like in Gaussian 09, as well as tighter two-electron
       integral accuracy. This can indeed cause some serious difference
       in results (UltraFine is much better). But if we compare HF with
       various basis sets, for small molecules second hyperpolarizability
       results were not different more than by 0.002 % for me.

_________

Yours sincerely,
Igors Mihailovs
Research assistant
Laboratory of Organic Materials
Institute of Solid State Physics
University of Latvia

www.cfi.lu.lv

On 04.03.22 00:30, Andrew DeYoung andrewdaviddeyoung##gmail.com wrote:

Hi,

My institution has both Gaussian 09 and Gaussian 16, and I'm trying to decide which to use for my new project.  (My background is in applications of MD simulations, and I'm quite new to quantum chemistry.)  One reason to use 16 is that, well, it's the latest version.  One reason to use 09 is for consistency; it's been used in our group a lot in the past.

My colleague said he observed some different results from 09 and 16, though he didn't remember how large the differences were.  

I will likely do some of my own comparisons, but I am curious, does anyone have experience with this?

The calculations I will be doing for my project are pretty mundane: using RHF to optimize geometry, and using MP2 to calculate the energy and extract charges with CHelpG.  The basis sets I will use are fairly pedestrian -- 6-31G(d), cc-pVTZ, etc -- nothing particularly exotic.  

From looking at this page ( https://gaussian.com/gdiffs/ ), it looks like Gaussian 16 mainly added new methods and basis sets.  I can imagine that algorithm improvements/tweaks (faster, more effective parallelization, etc) could give rise to some numerical differences, but I would not expect these differences to be terribly significant for most properties.  Is this your experience?

Thanks for your time,

Andrew

Andrew DeYoung, PhD

Department of Chemistry

Carnegie Mellon University