Learning Forum

[Nick Wallish]

We are exploring potential speed-ups using GPGPU for our flow studies.  Currently, we see no speed increase when using a fixed number of processor threads and with an Nvidia RTX A5000 GPU enabled or disabled. 

Originally we attempted this in v2022 R2, but updated to 2023 R1 to see if that would help.  Fluent is reporting that the GPU is there and selected, but we see nearly identical performance with/without the GPU.  This is not in Native GPU mode, but with the GPU added in the Parallel tab of the launcher.  Our case uses the following models: transient, VOF Multiphase, k-epsilon Realizable turbulence, and DPM.    I have attempted to modify the settings under /solve/set/amg-options/amg-gpgpu-options to enable for limited equations, and while after iteration it reports "AMG on GPU" in the console, there is still no speed change.

Is it possible to accelerate this simulation on GPU?

[Rob]

It's covered here   https://ansyshelp.ansys.com/account/Secured?returnurl=/Views/Secured/corp/v231/en/flu_ug/UsingGraphicsProcessingUnitsGPUsWit-EC38970B.html  but the gpgpu benefits are limited in many cases. 

Check how the load is balanced on the cpus. As VOF is a free surface tracker and DPM tends to have particles concentrated on some nodes you may find one or both of those models to slow the system down. Have a look at Parallel Load Balancing to see if that helps (with or without gpgpu). 

The native GPU solver will give you a perfomance improvement, but not (yet) with all physics. 

[Nick Wallish]

Hello, thank you for the response.  I have followed the instructions in that documentation, but that has only gotten me to where I was when I posted: enabling the GPU for some of the physics using the TUI options does not have ANY impact on performance.  To provide some numbers:  4 CPU and 0 GPU is approximately 10.9 sec/iter, 4 CPU and 1 GPU is 10.7 sec/iter.  32:0 results in 2.44 sec/iter and 32:1 results in 2.42 sec/iter.   These numbers imply the GPU is having zero effect on calculation speeds.

Am I missing something?  The documentation refers to the GPU working on coupled systems but not scalar systems unless enabled, but I see no benefit with or without enabling for scalar systems (implying it is also doing nothing for the coupled calculation).  Furthermore, I see no change in GPU resource usage through system monitoring (no change in memory usage, etc.).

Can you provide further insight into expected benefits and ensuring the GPU is being used?

[Rob]

With VOF you'll only have P-V coupled, which may not give you much of a boost in this case. It should be more noticeable with the single phase models with more coupled equations. 

[Nick Wallish]

Hello, I have an update:  After enabling pressure and momentum manually to use the GPU (and getting the message before iteration about AMG and compilation with CUDA runtime), I checked nvidia-smi to see if Fluent was using any GPU.  Next to both the fluent and cortex processes there is a n/a listed for GPU usage.  I verified that the GPU was selected as well.

So in essence: with GPGPU enabled, my simulation is not using the GPU at all, which explains why there was no speed change.  Can you provide some insight into why this might be? Are there further options to be checked?  What is the expected workflow or behavior when adding a GPGPU?

[Rob]

The gpu acceleration was only ever used on a few bits of multigrid and some of the radiation tracking schemes. Turn off VOF and see if that gives you any load on the gpu: the VOF part itself wasn't gpu compatible. Development have since moved onto the dedicated gpu solver, which gives significant acceleration for the physics it currently supports: further models will be available in time.