Thanks for the video! What are RAM memory requirements of multy gpu solver or does it rely on GPU memory only? OK, I figured that one out - obviously it uses GPU memory only.
A fluent native multi-GPU solver relies on both system RAM to load the data and GPU RAM to run the simulation. We advise you to get about 2Gb of RAM per 1E6 cells that you intend to run. So if you intend to run 8E6 cells, having 16Gb system RAM and 16GB GPU RAM should work well. Some specific physics/setup may require a bit more RAM but the recommendation above should work for most.
@@randsimulation Many thanks for your reply. One more thing - does NVlink technology have anything to do with multi gpu solver? I mean, is it helpful if using (for example) multiple V100 GPU's on one server? Also, where I can learn/read more about speeding up CoupledSolver DP with GPU? I understand that it is not as straight forward as with "native gpu solver". Information that I've found so far is very controversial and many people report worsening of performance when using GPU acceleration with CPU based solver.
Thanks for the video! What are RAM memory requirements of multy gpu solver or does it rely on GPU memory only? OK, I figured that one out - obviously it uses GPU memory only.
A fluent native multi-GPU solver relies on both system RAM to load the data and GPU RAM to run the simulation. We advise you to get about 2Gb of RAM per 1E6 cells that you intend to run. So if you intend to run 8E6 cells, having 16Gb system RAM and 16GB GPU RAM should work well. Some specific physics/setup may require a bit more RAM but the recommendation above should work for most.
@@randsimulation Many thanks for your reply. One more thing - does NVlink technology have anything to do with multi gpu solver? I mean, is it helpful if using (for example) multiple V100 GPU's on one server? Also, where I can learn/read more about speeding up CoupledSolver DP with GPU? I understand that it is not as straight forward as with "native gpu solver". Information that I've found so far is very controversial and many people report worsening of performance when using GPU acceleration with CPU based solver.