Has anyone else noticed that ORCA 6 is behaving slower than ORCA 5? For example, if I run the simple test job:
! B3LYP def2-QZVPP
%tddft
NRoots 5
end
*xyz 0 2
Ca 0.00 0.00 0.00
F 0.00 0.00 1.967
*
it completes in about 8 seconds on ORCA 5.0.4 but 14 seconds in ORCA 6.0.0. This is on a Mac M1 machine. I have run the identical input file under each version of ORCA, and the SCF iterations take about 2x longer in the ORCA 6 case.
Please note that in ORCA 6 we not only have performance improvements compared to ORCA 5, but also tightened many thresholds and increased the default grid size, so that the calculated results are more accurate. Generally speaking, increasing the grid size slows down all calculations by the same ratio, while our performance improvements (prescreening etc.) yield the greatest acceleration for large systems.
I've reproduced your calculations on a Linux node: the ORCA 5.0.4 calculation took 19.3 s and the ORCA 6.0.0 calculation took 29.6 s. The slowdown seems to be mostly due to a slowdown of COSX, as can be seen from the breakdown of SCF timings:
ORCA 5.0.4:
Total SCF time: 0 days 0 hours 0 min 14 sec
Total time .... 14.123 sec
Sum of individual times .... 13.985 sec ( 99.0%)
Fock matrix formation .... 13.377 sec ( 94.7%)
Split-RI-J .... 0.961 sec ( 7.2% of F)
Chain of spheres X .... 9.301 sec ( 69.5% of F)
XC integration .... 3.042 sec ( 22.7% of F)
Basis function eval. .... 0.684 sec ( 22.5% of XC)
Density eval. .... 0.871 sec ( 28.6% of XC)
XC-Functional eval. .... 0.230 sec ( 7.6% of XC)
XC-Potential eval. .... 1.175 sec ( 38.6% of XC)
Diagonalization .... 0.080 sec ( 0.6%)
Density matrix formation .... 0.004 sec ( 0.0%)
Population analysis .... 0.005 sec ( 0.0%)
Initial guess .... 0.122 sec ( 0.9%)
Orbital Transformation .... 0.000 sec ( 0.0%)
Orbital Orthonormalization .... 0.000 sec ( 0.0%)
DIIS solution .... 0.019 sec ( 0.1%)
Grid generation .... 0.376 sec ( 2.7%)
ORCA 6.0.0:
Total SCF time: 0 days 0 hours 0 min 19 sec
Total time .... 19.193 sec
Sum of individual times .... 19.180 sec ( 99.9%)
SCF preparation .... 0.071 sec ( 0.4%)
Fock matrix formation .... 18.985 sec ( 98.9%)
Startup .... 0.005 sec ( 0.0% of F)
Split-RI-J .... 0.638 sec ( 3.4% of F)
Chain of spheres X .... 15.574 sec ( 82.0% of F)
XC integration .... 2.745 sec ( 14.5% of F)
XC Preparation .... 0.000 sec ( 0.0% of XC)
Basis function eval. .... 0.689 sec ( 25.1% of XC)
Density eval. .... 0.635 sec ( 23.1% of XC)
XC-Functional eval. .... 0.165 sec ( 6.0% of XC)
XC-Potential eval. .... 1.142 sec ( 41.6% of XC)
Diagonalization .... 0.000 sec ( 0.0%)
Density matrix formation .... 0.012 sec ( 0.1%)
Total Energy calculation .... 0.003 sec ( 0.0%)
Population analysis .... 0.007 sec ( 0.0%)
Orbital Transformation .... 0.014 sec ( 0.1%)
Orbital Orthonormalization .... 0.000 sec ( 0.0%)
DIIS solution .... 0.063 sec ( 0.3%)
SOSCF solution .... 0.024 sec ( 0.1%)
Now let's have a look at the COSX grid. The GRIDX 1, GRIDX 2, and GRIDX 3 grids have 2050, 5002 and 11214 grid points in ORCA 5.0.4, while they have 7071, 8559 and 11214 grid points in ORCA 6.0.0. This means that ORCA 6.0.0 should give more accurate SCF wavefunctions and energies. You can see this from the energies of the last two SCF iterations; as ORCA uses GRIDX1 for the initial iterations, GRIDX2 for later iterations, and GRIDX3 for the last iteration, the energy difference of the last two steps is a measure of the quality of GRIDX2 compared to the large and accurate GRIDX3. The results are
ORCA 5.0.4:
second-to-last iteration: -777.4633964923, last iteration: -777.46342659
ORCA 6.0.0:
second-to-last iteration: -777.4634258690934985, last iteration: -777.46342608045552
That is, the GRIDX 2 of ORCA 5.0.4 has an error of 3e-5 Hartree, while that of ORCA 6.0.0 has an error of only 2e-7 Hartree. After recalculating the energy using GRIDX 3, the difference is reduced to 5e-7 Hartree, but it should be expected that the wavefunction is more sensitive to the quality of GRIDX 2, since the last iteration does not rediagonalize the Fock matrix. Indeed, the Mulliken populations of the two calculations differ by up to 7e-5, and the excitation energies differ by up to 0.001 eV.
To summarize, the slowdown of ORCA 6.0.0 for small systems is the price to pay for a higher accuracy, and for large systems ORCA 6.0.0 is expected to be faster AND more accurate than ORCA 5.0.4. For people who calculate both small and large systems, it's frequently the computational time of the large systems, and the accuracy of both the small and large systems, that matter.
