Advance GPU solver settings in EDEM

Sai Krishna_20256

Hello,

What are the following and their influence on simulation speed and accuracy of results?

1. Sphere-Geometry Grid- tried to use a custom grid cell, but when trying to change the value, it is always showing that the recommended option is selected. I am not sure if my custom size is applied or not.

2. Contact Detection options

3. GPU factory

 

Thanks,

Sai Krishna.

Stephen Cole

Hi Sai,

 

On point 1 we'll follow-up with this, I also see the same issue with the user interface always defaulting to "recommended". I've reported to our development team.

For the contact detection methods in EDEM we have the CPU solver and the GPU solver for sphere/multi-sphere particles, for Polyhedral particles there is a separate GPU solver.  For the sphere solver we have always tried to ensure the GPU and CPU solvers are identical, by default both solvers use the same contact detection method.

The default contact detection method was originally developed for CPU and is not always optimal for GPU, especially with large particle size distributions. When choosing the alternative contact detection method (ACD) this switches between algorithms depending on the simulation settings in order to find the optimal method, especially for large particle size distributions on the GPU. You can expect it to use less memory and is faster when there is a large size distribution (greater than 1:8).  The single precision option, which is also only available on GPU, runs all the calculations on single precision rather than double precision, this can be less accurate however most GPU cards are faster for single precision operations.  Overall any changes to the advanced settings for GPU will mean the GPU and CPU solvers are not identical.  When choosing ACD method this is no less accurate but may give different results to the CPU solver.

 

Factories are run on the CPU when considering the sphere solver. once the particle is created it remains processed on the GPU until it leaves the 'factory area' at which point it is passed over to the GPU.  As the CPU solver is slower than GPU ideally you want the factory area to be as small as possible (e.g. flat plane rather than volume) and for the particles to leave this area to be processed on the GPU.  

Also any particles that are processed on the GPU don't 'see' the particles in the factory, so if a GPU particle re-enters the factory then it can overlap and cause an explosion. This could occur if you have two factories one dropping particles on top of another. You can group the factories together to make a larger factory area to prevent this from happening.

 

There is some further information on this in the EDEM help > Simulator > GPU Advanced Settings.

 

Regards

Stephen