(03-10-2021, 06:31 PM)tyFlow Wrote: Yea what you're seeing is the IZ solver transferring over to CUDA at only a certain threshold, because up until that point it's faster to do small local collision solves on the CPU. Higher forces = more simultaneous collisions = greater chance of falling back onto CUDA.

The difficulty in debugging these issues is that they're hardware dependent....because I have systems with different GPUs/configurations that do not run into any trouble during solves. That said, I do have access to a 2080ti system and I will do further testing on it when I get a chance.

Anyways, thanks for posting your setup/info...hopefully we can track these issues down.

(03-10-2021, 07:30 PM)tyFlow Wrote: Could you guys confirm what your WDDM TDR delay setting is set to?

Instructions for how to find it are here if you have NSight installed:

https://docs.nvidia.com/gameworks/content/developertools/desktop/timeout_detection_recovery.htm

Or here if you don't:

https://www.pugetsystems.com/labs/hpc/Working-around-TDR-in-Windows-for-a-better-GPU-computing-experience-777/

Further instructions here:

https://docs.substance3d.com/spdoc/gpu-drivers-crash-with-long-computations-128745489.html

(03-10-2021, 07:42 PM)tyFlow Wrote: No, higher is better in this case

At least we've narrowed down that that's not the cause.

(03-11-2021, 05:08 AM)tyFlow Wrote: What version of windows are you guys running?

I tested on a machine with a 2080ti and could simulate all the CCCS example scenes without issue, with the latest NVidia drivers installed. However, it's a windows 8 machine...perhaps the problem is only happening on Windows 10?

(03-11-2021, 07:36 AM)frascow Wrote: Hi there,

My values on TdrDdiDelay and TdrDelay are already 60, Redshift renderer asked long time ago to add those lines to the registry,
About OS, Windos 10 (Buid 18363) pro and 3dsMax 2021

At least it seems that we both have exactly the same problem.

thanks!

(03-11-2021, 01:16 PM)JonY Wrote:

Quote:Frascow, you may want to see what my last post that discussed the observations and workaround I did in re to cuda crashing.

Hi JonY

I actually tested your workaround, setting the multiplier on a very low value and it works, unfortunately on the scenes I usually work the sense of scale and right physics is completely lost with this trick. Thanks for it but I really need a better solution, I'm sure Tyson will find it, meanwhile I prefer to work with the version 0.16108 that I keep running on 3dsmax 2020

__________________________________

I have attached a simple scene that generate the cuda error code 13 on my machine.

In the 3ds max scene, a simple cloth sticking to a box, and colliding with a box.

When the physical size of collider box is large, cuda solver works fine. But when the collider box becomes smaller (thinner), that is when cuda solver runs into problem - ie no collision after N frames. This correlate to your IZ solver point raised in your earlier post.

What I observed is rather straight forward:

. When collider object is large enough in size to collide with cloth, cuda solver works fine.

. When collider object becomes finer and intricate in size, cuda solver will crash after N frames. The way round this to what I tested was to play around with the Force parameters, by reducing the Multiplier value mentioned in my earlier post. This at least works for the moment.  


Anomaly observed in Point Cache (for suggestions):

. Upon a successful cuda solver cache of Tyflow with Mesh as its last operation (with Render off), I inserted a Point Cache to record the full simulation as individual cache files.
. But when I reload the cache on a copied plane from the original, either:
   . the new plane with loaded Point Cache disappears and not seen in viewport
or
   . the coordinates of the loaded Point Cache gets distorted, eg rotated and away from the original coordinates where cloth simulation was performed

---

(03-11-2021, 07:36 AM)frascow Wrote: Hi there,

My values on TdrDdiDelay and TdrDelay are already 60, Redshift renderer asked long time ago to add those lines to the registry,
About OS, Windos 10 (Buid 18363) pro and 3dsMax 2021

At least it seems that we both have exactly the same problem.

thanks!

Frascow, you may want to see what my last post that discussed the observations and workaround I did in re to cuda crashing.

(03-11-2021, 08:18 PM)frascow Wrote:

(03-11-2021, 01:16 PM)JonY Wrote:

Quote:Frascow, you may want to see what my last post that discussed the observations and workaround I did in re to cuda crashing.

Hi JonY

I actually tested your workaround, setting the multiplier on a very low value and it works, unfortunately on the scenes I usually work the sense of scale and right physics is completely lost with this trick. Thanks for it but I really need a better solution, I'm sure Tyson will find it, meanwhile I prefer to work with the version 0.16108 that I keep running on 3dsmax 2020

(03-24-2021, 08:57 AM)tyFlow Wrote: Hang tight for the next build....as of this evening a big CUDA bug happening on 2080ti (and newer) GPUs was found and fixed. Since I don't have a 30XX GPU to test I don't have a way of confirming that it fixes all the issues present on 30XXs too...but it fixes a pretty major one, and there's a possibility the bug in question being fixed will clear up some other issues too.

(03-24-2021, 08:57 AM)tyFlow Wrote: Hang tight for the next build....as of this evening a big CUDA bug happening on 2080ti (and newer) GPUs was found and fixed. Since I don't have a 30XX GPU to test I don't have a way of confirming that it fixes all the issues present on 30XXs too...but it fixes a pretty major one, and there's a possibility the bug in question being fixed will clear up some other issues too.