Performance
How fast simulations, analyses and exports run, and what to expect on other hardware.
All numbers on this page were measured on the development PC: one NVIDIA GeForce RTX 5070 Ti (compute capability 12.0, 16 GB, driver 610.88) with the CUDA 12.2 build of Geinium, an Intel Core i7-14700K processor (8 performance and 12 efficiency cores) and Windows 11. It is the only graphics card the GPU engine has been run on. On another card the speed will differ; how to measure it is at the end of this page.
Simulation speed
Conditions: mdsim bench -nb gpu, 20,000 timed steps of the complete NPT protocol
(thermostat, barostat, constraints, PME, energies every 100 steps), median of three runs, GPU
otherwise idle. The time for 100 ns follows from the speed.
| System and protocol | ns/day | 100 ns take | With multiple time stepping (option) |
|---|---|---|---|
| Lysozyme in water, 33,752 atoms, 2 fs (the default) | 1164.4 | 2 h 4 min | 1326.9 ns/day |
| The same, 4 fs with repartitioned hydrogen masses, two LINCS iterations (option) | 2134.9 | 1 h 7 min | 2274.0 ns/day |
| A protein-ligand system in water, 67,156 atoms, 2 fs (the default) | 674.9 | 3 h 33 min | 748.2 ns/day |
| The same, 4 fs with repartitioned hydrogen masses, one LINCS iteration in the input | 1246.4 | 1 h 56 min | 1342.0 ns/day |
These are the speeds of one replica. A job with three replicas takes three times as long when the replicas run one after the other; see below for running them side by side. Equilibration (200 ps) and minimisation add minutes, not hours. The pipeline writes two LINCS iterations at 4 fs, so a 4 fs job of the larger system runs somewhat slower than the last row. The output files of this release are byte-identical to those of release 1.0.0, which ran the default protocol 1.37 to 1.41 times slower: the gain came from scheduling, the pair search and the kernels, not from accuracy.
What affects the speed
| Factor | Effect | What you can do |
|---|---|---|
| Number of atoms | Roughly in proportion: twice the atoms, a little more than half the speed (674.9 against 1164.4 ns/day). | Use the smallest box that is safe: the default solute-box distance of 1.0 nm with a rhombic dodecahedron needs less water than a cube. |
| Time step | 4 fs with repartitioned hydrogen masses is about 1.8 times faster than 2 fs. | Choose 4 fs on the Run MD page when the faster sampling matters more than staying with standard masses. |
| Multiple time stepping | Long-range forces every second step: 1.11 to 1.14 times faster in the benchmark, 1.04 to 1.07 in 200 ps runs with all output files written. | An option for the production run (protocol.mts_factor: 2); it changes the integration scheme, so it is not the default. |
| Output | Writing frames and energies costs time; the benchmark writes energies every 100 steps. | The defaults (frames, energies and log every 10 ps) cost little. Avoid full-precision trajectories (trr_ps) unless needed. |
| Start-up | About 1.4 s outside the step loop at every start of the GPU engine; the first start on a new card or driver also compiles the GPU code once (6 to 8 s). | Matters only for very short runs. |
| Other programs on the GPU or processor | The engine raises the priority of the thread that drives the GPU, but a busy processor or another GPU program still slows it. | Leave the PC to the simulation for timing measurements. |
| The graphics card | Newer and larger cards are faster; the memory needed is small. | Run mdsim bench to see what your card does. |
Several runs on one GPU
A single run of a mid-sized protein does not fill a modern GPU completely, so two to four runs
side by side give more total throughput, while each single run becomes slower. Total throughput
relative to one run, measured with mdsim bench -batch 4:
| System | 2 runs | 3 runs | 4 runs |
|---|---|---|---|
| Lysozyme, 33,752 atoms | 1.10 | 1.15 | 1.18 |
| 67,156-atom protein-ligand system | 1.02 | ||
| Water box, 6,495 atoms (another session) | 1.55 | 1.86 | 2.06 |
Small systems gain much; large systems gain almost nothing. In a job, "Concurrent replicas" on
the Run MD page (project key replicas.concurrent) runs replicas side by side; "Automatic"
measures the best number once per job (about a second per candidate, at most 4 lanes) and writes the
result to pipeline.log. The results do not depend on it: replicas run side by side give
the same files as replicas run one after the other. The Batch queue runs one job at a time.
The CPU engine
The CPU engine is the double-precision reference implementation; it gives the same results but is
not a production engine. For 33,752 atoms on 8 threads it runs at about 6 to 9 ns/day, so 100 ns
would take 11 to 17 days. It is used automatically when no usable GPU exists, and always for energy
minimisation, which takes seconds to a minute (18 s in a 47,630-atom test job). On processors with
performance and efficiency cores, its force kernel is about twice as fast when the threads are pinned
one per performance core (engine.pin, mdrun -pin on).
Preparation, analyses, MM-GBSA and export
| Step | Time | Conditions |
|---|---|---|
| Structure, ligand, topology and system stages of a receptor complex | 16 s | A 61,705-atom system, with gap modelling |
| Modelling one missing loop | 1.5 s, 3 s, 20 s | Gaps of 2, 4 and 8 residues |
| Ligand parameters (AM1-BCC) | Seconds for most drug-like ligands; 69 s for the largest of the test set | 44 drug-like molecules took 451 s in total, on 8 threads of a shared PC |
| Interaction analysis of a 100 ns replica | 6 to 14 s | 10,001 frames, 47,630 to 89,031 atoms, 8 threads |
| Detail analyses of the results export | 41 to 63 s per 100 ns job; about 10 s for a thinned trajectory of 1001 frames | Run once per job; also from "Compute" on the Results page |
| MM-GBSA over 1,001 frames | 66 to 67 s (44 s with -fastexp) | 47,653 atoms per frame, 8 threads; the cost grows with the square of the solute size |
| A complete short job | 481 s for 23 stages | 3 replicas of 1 ns, 47,630 atoms, an earlier build |
GPU memory
The engine estimates the memory a run needs before it allocates anything: about 0.4 GB, plus about 2 kB per atom, plus 48 bytes per PME grid point. Measured: 207 to 307 MiB for systems of 6,495 to 88,884 atoms. If the free memory of the GPU is smaller, the run uses the CPU engine and says why.
Which GPUs are supported
| Requirement | Value | Notes |
|---|---|---|
| Graphics card | NVIDIA, compute capability 5.0 or newer (Maxwell, for example GeForce GTX 750 Ti and GTX 900 series, and every later generation) | The program contains ready machine code for compute capability 6.1, 7.5, 8.6 and 8.9 and portable code for everything from 5.0 on, which the driver compiles for newer cards. |
| Driver | 527.41 or newer | Cards that run the portable code may need a driver at least as new as the toolkit (536.25); not tested. |
| Detection | Automatic, at every start | Driver, cards, memory and the FFT library are checked; nothing to configure. |
| No usable GPU | The CPU engine is used | With a message that names the reason. The results are the same; the run takes much longer. |
mdsim gpuinfo -selftest before the first job.Measuring on your PC
rem which GPUs are there, and can the GPU engine use them?
mdsim gpuinfo -probe
rem does this GPU compute the same forces and energies as the CPU engine? (about two seconds)
mdsim gpuinfo -selftest
rem speed of a real run input, for example the production input of a finished job
mdsim bench -s P:\Geinium\Jobs\m3_dicyclomine\run\06_run\rep1\md.mdtpr -nb gpu
rem total throughput of 1 to 4 runs side by side, with a recommendation
mdsim bench -s P:\Geinium\Jobs\m3_dicyclomine\run\06_run\rep1\md.mdtpr -nb gpu -batch 4
bench runs 5000 timed steps by default (-nsteps) after a warm-up. For a
fair number, close other programs that use the GPU and run it two or three times.