MD Benchmarks for Amber, CHARMM and NAMD
(latest update: April, 2005)
Motivation for this benchmark
The benchmarking data below is intended to provide information useful
in assessing single processor performance and parallel scaling for
'typical' molecular dynamics simulations on a range of computer
architectures found in researchers laboratories and at National
Supercomputing Centers.
The data assembled here represents a snapshot of
performance on three widely used molecular dynamics programs currently
available to the scientific community: CHARMM, Amber and NAMD.
While
such information can be useful in guiding researchers as to which
programs are best suited for calculations on particular hardware
platforms, other factors including flexibility, functionality and
availability of a particular program enter into consideration as
well.
Our benchmark represents a
typical simulation of a small protein in water, of the type and
size that is often studied at the present time.
Input files are provided so that the timings can be reproduced and so
that other programs can be tested for the same MD parameters.
The most recent versions of the programs were used
for this benchmark in March, 2004:
Users might also be interested in looking at a version of this page from
2002; both Amber and NAMD have roughly doubled their speed in the past
two years, at least for some platforms.
Description of benchmark
The Joint Amber-Charmm (JAC) benchmark considers a protein (dihydrofolate reductase, dhfr)
in an explicit water bath with cubic periodic boundary conditions. Details of system size and simulation conditions are:
- 23,558 atoms
- protein: 159 residues, 2489 atoms
- water: 7023 molecules TIP3P, 21,069 atoms
- Cubic periodic box, 62.23 Å dimension
- 9Å nonbond cutoff with 2Å buffer, i.e., list with 11Å cutoff
- 1 fs timestep, 1000 steps
- NVE ensemble (constant energy, constant volume); CHARMM was run with NVT using Nose'-Hoover
- bonds to hydrogen constrained (SHAKE)
- PME electrostatics with 64x64x64 grid
- equilibration temperature was 300K
The input files for
all three programs can be found here.
For Blue Horizon, Amber 8 (pmemd) is the fastest of these four codes for
all parallel environments, whereas NAMD has the speed advantage on the more
modern DataStar machine. It is also clear that both Amber and NAMD have greatly
improved the speed of their codes over the past few years.
Blue Horizon at SDSC
(execution times in seconds)
| processors |
CHARMM
c31a2 |
Amber
8 |
Amber
7 |
Amber
6 |
NAMD
2.5 |
NAMD
2.4 |
| 1 | 2468 | 1708 | 2358 |
3116 | 1634 | 2986 |
| 2 | 1432 (1.7) | 891
(1.9) | 1229 | 1583 | 911 (1.8) | 1809 |
| 4 | 733 (3.4) | 459
(3.7) | 658 | 814 | 485 (3.4) |
930 |
| 8 | 440 (5.6) | 240
(7.1) | 375 | 442 | 272 (6.0) |
516 |
| 16 | 276 (8.9) | 128
(13.3) | 228 | 268 | 158 (10.3) |
308 |
| 32 | 221 (11.2) |
73 (23.4) | 162 | 190 | 95 (17.2) |
202 |
| 64 | 229 (10.8) | 47
(36.3) | 124 | 178 | 66 (24.8) |
124 |
|
Specifications:
|
Data Star at SDSC
(values in ps/day)
| processors |
CHARMM
31a2 |
Amber
8 |
NAMD
2.5 |
| 2 | 195 (2.0) | 257 (2.0) | 271 (2.0) |
| 4 | 368 (3.8) | 504 (3.9) | 533 (3.9) |
| 8 | 650 (6.7) | 952 (7.4) | 973 (7.2) |
| 16 | 993 (10.2) | 1799 (14.0) | 1918 (14.2) |
| 32 | 1271 (13.0) | 3138 (24.4) | 3431 (25.3) |
| 64 | 1252 (12.8) | 5310 (41.3) | 4906 (36.2) |
| 128 | | 6101 (47.5) | 8136 (60.0) |
|
Specifications:
|
Notes: The Amber 8 pmemd has not been optimzed for this machine,
so that better results might be possible; we don't know whether the same is
true for NAMD or not. The NAMD 64-cpu result seems out of line; we are
investigating this. Unlike the Blue Horizon case, the NAMD results are almost
always better than the Amber numbers, and this is especially true at large
processor counts. Speedups (in parenthesis) assume that two processors has a
speedup of 2.0; this may slightly overestimate the actual parallel efficiency.
The configuration file for pmemd on Data Star is
here.