GPU memory question
Hello everyone,
I am sorry if this is a stupid question. So I got a RTX3090, and I am trying to simulate a high resolution dam-break case. I am using the precompiled Windows-version of DualSPHhysics with the latest Nvidia drivers (531.18 Nvidia game ready driver/528.49 Nvidia studio driver).
The GPU has a total memory of 32 GB. However, if I am trying to decrease the particle distance to about 2.5 mm, the solver is giving me a: "Failed GPU memory allocation. (CUDA error 2 (out of memory))". Nvidiainspector and the Windows task manager are showing me a memory usage of about 10 GB for a simulation with a particle distance of 3 mm. Therefore, I am assuming that a marginally smaller particle distance should also work. However, if I decrease the particle distance slightly (e.g. to 2.9 mm) I am also given the error from above. CUDA is showing a usage of about 95% in the Windows task manager for a simulation with 3 mm.
So I do not understand why I cannot decrease the particle distance by a greater amount. Am I misunderstanding the usage of memory here? Is there a different memory limit for CUDA? Is there some special Windows setting to use all the memory? I really do not understand why the Solver is complaining, as I can use the full memory on my laptop with a Linux environment and a GeForce RTX 3050 Laptop.
Thanks for your help in advance
Felix
Comments
Hi Felix, actually I have tried several times in Tesla GPU card[80 Gb memory], I find the same issue when the total particles is nearly 180 million, i still don't understand. i ever tried the cpu version and find the same issue when total particles is close to 180 million in one simulation case, i thought maybe there is some limitaions with the released verion, or maybe i missed something. see the attached thread https://forums.dual.sphysics.org/discussion/comment/5612#Comment_5612.
@imartinezestevez and @jmdalonso can help here
@yueguotian I do not think it is the same problem for me as my graphics card has much lower VRAM. So I looked around the web for a bit and found this thread in the old Microsoft tech support:
I was wondering whether somebody else has a workstation with Windows 10 available who could confirm that windows is allocating and blocking at least 20% of the VRAM for single processes. For me, it also seemed to be RAM-usage dependent, i.e. higher RAM-usage led to a lower amount of possible memory allocation. Therefore, I cannot use the full potential of the graphics card due to Windows 10? It might answer my question and would actually be a huge deal for running DualSPH on Windows 10.