Hi everybody,
I was wondering if lammps can be compiled for use with any graphics card (GPU), or only certain cards are supported. For example, I know of benchmarking runs of lammps with the Nvidia Tesla GPU on Nvidia's website, but I could not confirm if it will be possible to run lammps on older cards like the Nvidia Quadro.
Before buying a GPU, is there a way to confirm if it will support lammps? Does lammps have any minimum requirements for the GPU hardware on which can run?
Thanks
Steve
Hi everybody,
I was wondering if lammps can be compiled for use with any graphics card (GPU), or only certain cards are supported. For example, I know of benchmarking runs of lammps with the Nvidia Tesla GPU on Nvidia's website, but I could not confirm if it will be possible to run lammps on older cards like the Nvidia Quadro.
the GPU module in LAMMPS currently _only_ support CUDA which in turn _only_
works with Nvidia GPUs. there are several generations of nvidia GPUs
that support
CUDA to a different degree.
you can find out which nvidia GPUs support CUDA here:
http://developer.nvidia.com/cuda-gpus
Before buying a GPU, is there a way to confirm if it will support lammps? Does lammps have any minimum requirements for the GPU hardware on which can run?
having a GPU with at least compute capability 1.3 is highly recommended.
also, you should note that _both_ the number and clock of GPU cores as
well as memory clock and buswidth have a significant impact on the amount
of GPU acceleration.
for testing and cost efficiency reasons, it is probably best to purchase
a geforce class GPUs. out of the current models, a GeForce GTX 550Ti
is probably the minimum that makes sense for testing. the GPU in this
card has only 1/4 the cores of a GeForce GTX 580 which is probably the
fastest single chip GPU available currently (with 512 cores). Tesla family
GPUs have _much_ more memory, the option to enable ECC error checking
on memory, and 4x more double precision floating point units than geforce
family cards. this will be of advantage when compiling the GPU package
in full double precision mode. for mixed or single precision mode, this is
not the case.
cheers,
axel.