Scaling and New Functional Testing

go back to Main Page, Computational Resources, Chemistry & More, Computational Codes, GAUSSIAN, G09

Tests calculations performed at the CESCA supercomputer on parallel8 queue (prades) with 14000MB ram.

Calcfc, opt to TS, and frequency.

Job has 687 basis functions with B3LYP method.

     Step	           G03	          G09
calcfc to 2nd l103	7h 54m 08s     3h 42m 10s
2nd opt step	            8m 00s	   9m 16s
3rd opt step	            6m 06s	   7m 06s
freq	                8h 30m 25s     4h 06m 15s
Total calc time        16h 38m 39s     8h 04m 47s

So the frequency calculation is about 2x faster in Gaussian09.

The SCF takes longer, but it is using the new GEDIIS algorithm (vs GDIIS in G03 which now doesn't exist). They say GEDIIS should give better performance, especially for calculations not so close to convergence as this example.

Testing new functionals with G09

Job has 687 basis functions, starting structure from B3LYP method in G03.

     Step                          Functional		
   measured             M06           M06L           B97D
calcfc to 1st l103   4h 49m 34s     3h 08m 50s     1h 59m 42s
2nd opt step            13m 36s        10m 32s         7m 42s
3rd opt step            13m 37s        11m 10s         7m 15s
Total opt time      19h 05m 32s     7h 19m 06s     3h 36m 35s
No. steps                   89             44             32
avg opt step time          772s           599s           406s
freq                 5h 27m 59s     3h 15m 04s     2h 11m 48s
Total calc time     29h 50m 18s    14h 04m 42s      7h 48m 5s

The M06 functional is 33% slower than B3LYP (based on the frequency times), but still faster than B3LYP in G03. M06L is slightly faster than B3LYP. Grimme's B97D functional is very fast, and also takes less steps to optimize in this case too.