Recently, a variety of groups around the world formed the "ABC" consortium, in order to carry out collectively a large range of DNA simulations. The first round of simulations looked at 15-base pair segements of DNA with various sequences. This large scale effort looked at 39 oligomers, each with about 24,000 atoms, with each sequence simulated for 15 ns.
One of the sequences I had was "GACT". I used this input file for some benchmark tests on various processors, and with various versions of Amber. The ABC simulations were carried out with either Amber 6 or Amber 7, since Amber 8 was not yet released when the calculations were done. My tests did a 5 ps run, then multipled the results by 3000, to estimate the time needed to do a single 15 ns "run". Results are expressed in days.
Amber6 Amber7 Amber8 Amber8 Amber9
sander sander sander pmemd sander/ips
==========================================================================
SGI Altix, 8 cpu 16.5 8.8 7.3 6.4 (4.7)
SGI Altix, 16 cpu 11.8 6.5 5.1 4.2
Xeon, 2.0 GHz, 2 cpu 86.3 77.4 64.6 60.8
===========================================================================
The SGI Altix is bulit on 1.3 GHz Itanium II chips. If you were fortunate enough to have access to 8 processors on such a machine, you could run one of the ABC sequences in less than a week. At the other end of the spectrum, last year's (2 GHz) dual pentium machine would require about two months for the same job.
In all cases, it pays to upgrade to more recent versions of the Amber code, especially if you are still using Amber 6.
This test case is almost exactly the same size as the "jac" (dihydrofolate reductase) benchmark, for which Amber 8 benchmarks and comparisons between Amber, CHARMM and NAMD are available. (Note, however, that the ABC simulation used a time step of 2 fs, whereas JAC uses 1 fs; hence 15 ns of ABC is roughly as costly as 7.5 ns of JAC.)
Note that sander/ips is actually a different approximation (Isotropic Periodic Sums), so that the comparison is flawed. The general results with IPS are (surprisingly?) close to those from PME, but we have also noticed some systematic differences between the two. It is not yet clear how significant these are. Hence, the IPS number is in parentheses for a good reason.