Endonuclease PvuII (1PVI) DNA - GATTACAGATTACA
CAP - Catabolite gene Activating Protein (1BER)
DNA - GATTACAGATTACAGATTACA Endonuclease PvuII bound to palindromic DNA recognition site CAGCTG (1PVI) DNA - GATTACAGATTACAGATTACA TBP - TATA box Binding Protein (1C9B)
CAP - Catabolite gene Activating Protein (1BER)
GCN4 - leucine zipper transcription factor bound to palindromic DNA recognition site ATGAC(G)TCAT (1YSA)
GCN4 - leucine zipper transcription factor bound to palindromic DNA recognition site ATGAC(G)TCAT (1YSA)
GCN4 - leucine zipper transcription factor bound to palindromic DNA recognition site ATGAC(G)TCAT (1YSA)
GCN4 - leucine zipper transcription factor bound to palindromic DNA recognition site ATGAC(G)TCAT (1YSA)
GCN4 - leucine zipper transcription factor bound to palindromic DNA recognition site ATGAC(G)TCAT (1YSA)
GCN4 - leucine zipper transcription factor bound to palindromic DNA recognition site ATGAC(G)TCAT (1YSA)
TBP - TATA box Binding Protein (1C9B)
 

The YASARA Benchmarks

Since the first YASARA website went online in 1999, we have been providing benchmarks to allow an objective comparison with other programs. Our goal is however not to start 'benchmark wars' with other groups, as we are all trying to solve the same problems. So instead of comparing other programs with YASARA directly, we offer symbolic rewards of 100$ for the first one who can beat YASARA in one of the following benchmarks. (You do not have to use your own program, just anything you find).

The following restrictions apply:

  • Benchmarks are shown for one or more combinations of CPU model, RAM and graphics card. Your program must be faster on the same or a slower PC.
  • Your program must be available, you must be able to reproduce your results.
  • Only the first one who wins a benchmark gets the reward, then the benchmark will be removed.

Benchmark 1: 100$ for a faster molecular dynamics algorithm

Dihydrofolate ReductaseThe image on the right shows a classic molecular dynamics benchmark: dihydrofolate reductase (PDB entry 1AOE) and 6875 water molecules, summing up to 23788 atoms. You can download the PDB file here. Here we measure the time required for one full simulation step of a standard MD protocol using the AMBER03 force field and TIP3P water:

  • Find the neighbors within a 7.86 Å cutoff distance (usually via a grid-based search).
  • Calculate the intramolecular (bonds, angles, dihedrals) and shortrange intermolecular (Coulomb, VdW) forces, including the PME shortrange damping function.
  • Calculate the longrange electrostatic forces using the PME algorithm: interpolate the charges on a grid with <1 Å spacing and 4th order B-splines, perform the two fast fourier transforms and convolution, subtract the longrange component from bonded intraaction partners.
  • Integrate the equations of motion and move the atoms one step further.

Here are the execution times on various CPU types (single-core only, see next benchmark for multi-core performance). In our hands, especially the SSE4.1 code path is about 60% faster than the closest competitor (measured in 32bit Linux with standard GCC compiler).

CPU model
Opteron 265
Core 2 Duo
Xeon L5420
Code path
3DNow! + SSE + SSE2 + SSE3
SSE + SSE2 + (S)SSE3 SSE + SSE2 + (S)SSE3 + SSE 4.1
CPU cycles
292.91 million
231.78 million
189.95 million
Actual time
0.1627 seconds @ 1.8 GHz
0.1073 seconds @ 2.16 GHz
0.07598 seconds @ 2.5 GHz


Notes: How the results above translate to simulated nanoseconds per day depends on the chosen timestep and thus the accuracy/performance tradeoff you are willing to take. If bonds to hydrogen atoms are constrained and rigid water molecules are used, YASARA reliably reaches a timestep of 4 femtoseconds.

Benchmark 2: 100$ for a faster multi-core molecular dynamics algorithm

Dihydrofolate ReductaseToday, even entry level CPUs contain more than one core. While the single-core performance in the previous benchmark is a good measure for the degree of optimization and algorithmic tuning, in practice the multi-core performance is more important. The system simulated here is the same as in the previous benchmark, but multiple CPU cores are used now. The numbers below are for static load balancing, i.e. if you run the same simulation a second time, you will get the same trajectory. This is very important for several, but not all applications of MD. If your program can beat YASARA only with dynamic load balancing (and thus non-reproducible trajectories, covering a smaller range of applications), the reward is cut to 50$.

Here are the execution times on a growing number of CPU cores:

CPU model
Opteron 265, 1 core
Dual Core Opteron 265, 2 cores
2 x Dual Core Opteron 265, 4 cores
Code path
3DNow! + SSE + SSE2 + SSE3
3DNow! + SSE + SSE2 + SSE3 3DNow! + SSE + SSE2 + SSE3
CPU cycles
292.91 million
153.77 million
78.64 million
Actual time
0.1627 seconds @ 1.8 GHz
0.0854 seconds @ 1.8 GHz
0.0437 seconds @ 1.8 GHz


Notes: How the results above translate to simulated nanoseconds per day depends on the chosen timestep and thus the accuracy/performance tradeoff you are willing to take. If bonds to hydrogen atoms are constrained and rigid water molecules are used, YASARA reliably reaches a timestep of 4 femtoseconds.

Benchmark 3: 100$ for a more accurate force field

2A0B CrystalOn the left, you can see the unit cell of PDB entry 2A0B: the phosphotransfer domain of anaerobic sensor kinase ARCB, solved at 1.57 Å resolution. The cell contains four chains, and thanks to the high resolution, provides a very accurate view of protein structure. Running a molecular dynamics simulation of such a crystal moves the protein away from where it should be, due to inaccuracies in the force field. The smaller the damage, the more accurate the force field. During a 1 nanosecond simulation with YASARA's YAMBER2 force field, the maximum Calpha RMSD from the true structure is 0.70 Å. 100$ are yours if you manage to obtain a lower RMSD with any MD program / force field you like, under the condition that you did not optimize the force field specifically for 2A0B or its homologues.

Notes:
Hydrogens have been added with WHAT IF's H-bond network optimizer, counter ions and disordered water molecules were placed by YASARA. The cell size is 30.456 x 34.924 x 110.741 Å, you can download a PDB file here . Simulations must be run at the temperature of experimental structure determination (277K). Save snapshots in intervals of 5ps and calculate the average Calpha RMSD for all four chains. This average must stay below 0.70 Å in all 200 snapshots to win the benchmark.


Benchmark 4: 100$ for better molecular graphics

Ribosome, all atoms
Ribosome, zoomed in

The screenshots on the left show PDB entry 1N8R, the 50S ribosomal subunit with 98569 atoms, in space-filling mode. In the upper image, all atoms are on screen, YASARA delivers either 10 frames per second without shadows (many programs are a factor 35 slower), or 5.5 frames per second with shadows calculated in real-time (as shown). On the lower image, we flew right through the ribosome to the backside, most atoms are now off-screen, resulting in 106 frames per second without shadows (as shown). To win the 100$, your program must be faster in both cases (either with or without shadows), while reaching at least the same visual quality (e.g. spheres must be equally round), and it must run on the same or a slower system.

Notes:
Benchmark system:
Opteron 265 single-core @ 1.8 GHz, Window size 1024x768, color depth 32 bits, nVIDIA Geforce 6600 graphics card.


Benchmark 5: 100$ for faster interactive molecular dynamics

Interactive MD of Dihydrofolate ReductaseInteractive molecular modeling during a real-time molecular dynamics simulation allows you to pull individual atom and entire molecules around with the mouse or 3D input devices. The image on the right shows the simulation of dihydrofolate reductase (see benchmarks 1 and 2 above for details), including all water molecules. Using four CPU cores on the system described below, YASARA reaches 12 frames per second, i.e. 12 molecular dynamics steps and 12 screen updates including real-time shadows. To win the 100$, your program must be faster on the same or a slower system and reach at least the same visual quality.


Notes: Benchmark system: Opteron 265, four cores @ 1.8 GHz, Window size 1024x768, color depth 32 bits, nVIDIA Geforce 6600 graphics card.