Learning Forum

[papp]

 

Hi,

I've detected a strange behavior of the newest version of Discovery Live (2020R1). The latest release utilizes what's called Smart Memory Usage; however, the minimum feature size is significantly smaller compared to the previous version (2019R3).

I've tested a 1x1x1 m cube with the highest Fidelity setting on a GTX 1080Ti with 11 GB VRAM, the results are the following (see the attached image for more):

Version	Min. feature size	Cell count
2019 R3	4.7 mm	9 631 777
2020 R1	7 mm	2 915 452

Based on this experiment, either the spatial resolution is decreased substantially, or the mesh size is calculated erroneously by the software. The issue is present for different geometries and GPUs as well.

I kindly ask you to look into this problem! Also, please confirm or deny whether the minimum feature size is equal to the mesh size or not.

If possible, please share some details with us about how Smart Memory Usage works as well. (How the cells are divided around the thin objects, what's the condition of being thin, how far from the wall the refinement is present, etc.)

Thanks,

Bálint

[Subashni Ravichandran]

Hello Bálint Papp

I'm investigating this issue from my end. I will get back to you with some observations soon.

[papp]

Hi Subashni

You might want to take a look at the VRAM usage of the current and previous versions. On my computer at maximum Fidelity, Live 2019R3 uses 5.7 GB (which is surprisingly low), while Live 2020R1 uses only 3.2 GB out of the available 11 GB.

Can you experience something similar?

[Subashni Ravichandran]

Hello Bálint Papp

As per the Discovery Live 2020 R1 Update document   "New memory-efficient discretization approach that enables simulation of thinner parts and the use of smaller GPUs" has been implemented.

This means there has been significant changes in GPU usage to accommodate thin structures and use of smaller GPUs.

Also, it is a known fact that Discovery Live does not use the Graphics memory completely as it reserves some memory for geometry(or other) modification that may be made later on or for other applications in the system that may require it so the memory usage may improve or change based on modifications made.

No other information regarding the meshing technique has been shared.

🛈 This post originally contained file attachments which have been removed in compliance with the updated Ansys Learning Forum Terms & Conditions

[papp]

Hello Subashni

I'm very well aware of the goal of the new discretization method, and I completely understand if you are not allowed to share any further details about how the mesh is set up.
Thank you for clarifying the VRAM management situation. But please help me understand the behavior of 'Smart Memory Usage', because I'm confused knowing the results.

The minimum feature size is significantly larger in Live 2020R1, and hence, in the original example, there are roughly 3.3 times fewer cells using the same hardware and Fidelity compared to Live 2019R3.
Using the dump truck sample model, the difference reduces to around 2.9 times fewer cells, probably because the truck has more 'thin' features than the original cuboid body. (Minimum feature sizes: 2019R3: 47 mm, 2020R1: 67 mm; GPU: GTX 1080Ti, 11 GB VRAM; Fidelity: 100%.)

Please answer my question: is this behavior normal? In other words: can we expect to have a coarser mesh in 2020R1 compared to 2019R3, if there are no thin geometries in the domain?

Thanks,

Bálint

[Subashni Ravichandran]

Hello Bálint Papp

Unfortunately, this is the expected behaviour in 2020R1. There are some complications involved which are developers are trying to rectify to enable better capturing in the future versions. 

[papp]

Subashni OK, thanks for your answer.