Dear Developers and Users
I’m going to buy a GPU to accelerate my simulations. In this way, GTX 1080 Ti is the fastest GPU (of course in the Geforce series). However, I’ve just come across the Titan V, which is said to be majestically very fast. But, I’m afraid of that kind of GPU due to the following concerns:
(1) Titan V is developed based on Volta architecture. Could LAMMPS be compiled upon that architecture on a GTX card?
(2) The NVIDIA announced Titan V being of extraordinary 110 TFLOPS calculating power? That 's a practical statement?
(3) Lastly, I’m not sure if there are other determining factors other than TFLOPS to judge over GTX card upon having other items equal, such as CPU or RAM.
Any hint is welcome.
Quick answers on your questions:
- Yes, you can compile to any architecture using the NVIDIA compiler nvcc.
- No. They report 110 TFLOPS, but on half precision which is not useful for molecular dynamics simulations. See https://images.nvidia.com/content/technologies/volta/pdf/tesla-volta-v100-datasheet-letter-fnl-web.pdf. It does, however, have ~14 TFLOPS single precision and ~7 TFLOPS double precision.
- If you’re running simulations using the KOKKOS package and all of your fixes / computes run on the GPU, then I don’t think it matters too much what CPU/memory you have. If you use computes/fixes that run on CPU, or you are using the GPU package, the CPU clock speed (I’ve measured different cpu’s, and clock speed seems to be important) will affect the performance greatly.