GeinDFT
Home › Performance

Performance

How long calculations take, how they scale across processor cores, and how much memory they need.

The figures below were measured on an Intel Core i7-14700K with 32 gigabytes of memory, running Windows 11. Each is a wall time for the complete job, not for one part of it.

Typical calculations

JobMoleculeMethod and basisWall time
Single point with properties and surfacesWater, 3 atomsB3LYP / def2-TZVP2.6 s
Single point with mapped potential, orbitals and densityAspirin, 21 atomsB3LYP / def2-SVP25 s
Single point with mapped potentialCaffeine, 24 atomsB3LYP / def2-SVP24 s
Optimisation and frequenciesWater, 3 atomsB3LYP / def2-SVP8.9 s

Scaling across cores

Job20 threads8 performance cores
Aspirin, single point8.7 s11.0 s
Aspirin, geometry optimisation159 s212 s

Parallel efficiency on the dominant steps is between 94 and 98 per cent at eight cores. Going beyond eight on a consumer processor gains less than the core count suggests, because the additional cores on such chips are the slower efficiency cores: eight of those take about twice as long as eight performance cores on the same work. An ordinary eight core desktop is therefore a perfectly reasonable machine for this program.

Memory

JobPeak memory
Aspirin, single point302 MB
Aspirin, optimisation428 MB
58 atoms, 632 basis functions3.9 GB

The program estimates the memory a job will need before it starts and refuses a job that would exceed the limit, rather than starting it and exhausting the machine. On a 16 gigabyte machine the practical ceiling with the default settings is around 70 atoms of that kind at a double-zeta basis.

The window stays responsive

The calculation runs on worker threads. During a complete aspirin job the viewer kept drawing at about 50 frames per second, so the structure can be rotated and measured while the calculation proceeds.

The dispersion correction is free in practice

System8 threads20 threads
192 atoms0.46 ms0.39 ms
375 atoms0.99 ms0.93 ms
5184 atoms57 ms40 ms

Those times are for the energy and the analytic forces together, with every atom pair included rather than a distance cutoff. At that cost there is no reason to approximate it.