Our colleague suggests that, by bought the $1000 VGA card to our computer, our simulation would run incredibly faster, and can easily achieve 100 ns MD simulation in less than a week even though our computer has the standard specification (8GB RAM, Intel i5 processor, 500 GB HDD). Assuming our MB and PS could support the VGA card specification, should we buy it only instead of improving the other parts ? such buy the SSD or another 8GB RAM ?
Performance depends on the size of the system and the software being used. I know nothing about docking on GPU, but as for MD, even cheaper GPU can be used to achieve that kind of performance, with the right software. OpenMM is a pure GPU implementation and we can get anywhere from 20-150 ns/day depending on the size of the system. GROMACS performs similarly, but requires balanced CPU resources (see, e.g. http://dx.doi.org/10.1002/jcc.24030).
Performance depends on the size of the system and the software being used. I know nothing about docking on GPU, but as for MD, even cheaper GPU can be used to achieve that kind of performance, with the right software. OpenMM is a pure GPU implementation and we can get anywhere from 20-150 ns/day depending on the size of the system. GROMACS performs similarly, but requires balanced CPU resources (see, e.g. http://dx.doi.org/10.1002/jcc.24030).
As Justin Lemkul pointed out, accelerating MD simulations with GPUs depends on the software you are using for MD and how it uses the GPU and/or the CPU for computations. It also depends on the size and complexity of the system being simulated. If the code uses CUDA cores, you need to have a CUDA-capable video card with the CUDA drivers installed. Some MD software uses a combination of CPU and GPU computing using OpenCL instead of CUDA. Our group has used CPU/GPU computing, and a combination of an i7-5960 CPU running at 4.4 GHz and an Nvidia GTX 980 in a linux workstation enabled a DHFR/explicit solvent benchmark of 250 ns/day.
Rudy: could you give more info about your i7-5960 setup (particularly the amount of RAM memory and storage (SSD or HD?)) and what software did you use to get 250 ns/day? ? Did you do any tests with smaller/larger amounts of RAM?
Bartosz, my i7 workstation has 32 GB of RAM and 2 x 1 TB SSDs (Samsung 850 Pro). It is a custom-built computer from Puget Systems.
However, the main determinants of its performance in MD simulations are the CPU clock speed (overclocked at 4.4 GHz) and the GTX 980 GPU.
The software I am using currently for MD simulations is YASARA-Structure.
As Rudy and Justin pointed out, it depends on what types of systems you are working with and what your software package is. If you're using AMBER, you can have a look at http://ambermd.org/gpus/recommended_hardware.htm for an idea of what types of GPUs to buy. AMBER only has support for NVIDIA right now. For $580 USD you can get a GTX 780, which has been shown on a test system to achieve 250 ns/day (http://ambermd.org/gpus/)
Can you provide us with more information? What MD simulation software package do you typically use?
Let me add my experience about MD on a rather high-end GPU system. Our group at Delft University bought a Dell C4130 quad-K80 system. It cost quite a bit, something like €25k, but performance is awesome. We use it for doing lammps MD calculations with the eam interaction model (metals). For a calculation that fills up the 96 GB total VRAM on the 4 K80 cards (50-60 million atoms) the speedup over using only the 32 cpu cores in that machine can exceed a factor of 10, depending on the number of neighbours within the cut-off range. So a pretty compact 1U machine can give the computing power of over 300 Xeon cpu cores.
According to Dell's documentation for that machine (file attached) lammps is one of the best performing codes on gpus, so your experience with other MD codes may vary.
I have no connection to Dell btw. As an old-time Mac user, I may actually nourish some wary sentiments about Dell, but I have to say their C4130 hardware, while far from cheap, is terrific.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics