LCCSD(T) accuracy vs speed

Are there some benchmark data available for the different local-CC modes, i.e. localcc=2013/5/6/8 and for medium to larger molecules?

Dear chburger,

thank you for your interest in our code.

We are working on two reports focusing on the accuracy and efficiency of LNO-CCSD(T).
The first, Ref. 35 of the manual, was just submitted a few weeks ago.
There are also some limited data in Ref. 29 and 32 of the manual, but those are for the 2016 version.

I recommend to use the default 2018 version, that is clearly the fastest and most accurate.
The 2013/15/16 options are temporarily kept so that results of previous papers can be reproduced.

The 2016 version of local MP2 is well documented in Ref. 28. The new defaults for the 2018 LMP2 are almost identical to the tighter options in Ref. 28.
Every example of that LMP2 paper is feasible with the Normal settings of the 2018 LNO-CCSD(T). The Normal LNO-CCSD(T) will take about 5-10 times longer than the LMP2 and currently requires about 2-3 times more memory. The Tight LNO-CCSD(T) is about 4-5 times more expensive than the Normal. (Assuming that about 1-200 GBs of local hard disk is also available.)

The local error of LNO-CCSD(T) with the Normal settings is very small compared to the basis set incompleteness error at the TZ level.
If aiming for chemical accuracy close to the CBS limit the Tight option might be better.
It is useful to look at the convergence of the Loose-Normal-Tight results on at least one representative example.

For closed shell, 3D organic molecules I performed Tight-LNO-CCSD(T) with augmented QZ basis up to about 50-100 atoms in a few weeks on 8 cores. The corresponding basis set extrapolated reaction energies are expected to be within 1 kcal/mol of CCSD(T)/CBS.

It is hard to say more in general, the accuracy and efficiency varies with the system and the quantity of interest.
If it helps I am happy to assist you further to estimate the computational requirements for your system if you can specify it and your hardware more precisely.

All the best,
Peter

Just letting you know that our first report on the performance of the 2018 LNO-CCSD(T) is appeared online today :
Optimization of the Linear-Scaling Local Natural Orbital CCSD(T) Method: Improved Algorithm and Benchmark Applications
J. Chem. Theory Comput., DOI: 10.1021/acs.jctc.8b00442
pubs.acs.org/doi/abs/10.1021/acs.jctc.8b00442

A second paper is being written including more benchmarks and efficient protocols to approach the CBS limit of CCSD(T) with LNO-based methods.

I hope this helps.
I am happy to assist you further if you have more specific questions,
Peter