Table 2.
A summary of different run options is given and their effect on the observed energy drift. A “*” in the Lookup column indicates that the run was done on GTS250. The symbol “**” is for runs on the GTX470. The rest of the runs were conducted on GTX480. The energy drift is expressed as percentage of the total change in a microsecond. Quad-4096 is a table lookup for non-bonded interactions with 4096 entries, Lin-8000 is a lookup table that is using linear interpolation, Quad-256 is a table with 256 entries that is using quadratic interpolation. The PME calculation is done either in double precision (DP), or in single precision (SP). The cut provides the two cutoff distances that are used in MOIL. The NB Force option is the use of single precision or double precision in adding the non-bonded interactions. SHAKE B/L are the options of SHAKING all bonds (B) or just light particles (L), which usually means hydrogen atoms. The tolerance of the SHAKE algorithm is fixed at 10-12 relative error and the reciprocal sum of PME at 10-9. Note that the first row in the table is the most accurate option and the observed drift is within the noise level. Note also the horrible drift seen with 1fs time step and without the application of SHAKE (all water molecules are kept rigid with a Matrix SHAKE algorithm). ACEMD results are taken from reference8 for 1fs time step. Note that we tried two grid sizes for the reciprocal sum calculations, grid size of 32 and 64 points at each Cartesian direction. The effect on energy conservation was very small. Grid size impacts, however, pair correlation functions and appropriate choice must be made for the specific system at hand. See text for more details.
| NB Force | ShakeB/L | DRIFT | |||
|---|---|---|---|---|---|
| Lookup | PME | Cut (box-calc) | add | (tol 10-12) | (%microsec) |
| Quad-4096 | DP64 | 10.37-9.5 | DP | B | 0.016 |
| Quad-4096 | DP64 | 9.8-9.0 | DP | L | 0.053 |
| *Quad-4096 | DP64 | 10.37-9.5 | SP | B | 0.27 |
| Quad-4096 | DP64 | 8.89-8.5 | SP | B | 0.52 |
| Lin-8000 | DP64 | 10.37-9.5 | SP | B | 0.82 |
| *Quad-4096 | SP64 | 10.37-9.5 | SP | B | 3.2 |
| Quad-256 | SP32 | 8.89-8.5 | SP | L | 2.4 |
| Quad-256 | DP32 | 8.89-8.5 | SP | L | 0.84 |
| **Quad-256 | DP32 | 8.89-8.5 | SP | L | -0.60 |
| **Quad-256 | SP52 | 8.89-8.5 | SP | L | 3.57 |
| **Quad-256 | DP52 | 8.89-8.5 | SP | L | -0.30 |
| Lin-8000 | SP32 | N/A-8.9 | SP | NO | 40. |
| Quad-4096 | SP32 | 8.89-8.5 | SP | NO | 43.0 |
| *Quad-4096 | DP64 | 9.8-9.0 | SP | NO | 32.0 |
| ACEMD | NO | 10.15 |