GPU Accelerated Moldlfow Simulation

[PAVITHARAN_V]

Dear Autodesk Team,

   Recently few CFD Solvers (namely few StarCCM+ & Ansys Fluent) have seen a tremendous increase in their performance through the implementation of GPU Acceleration in their program so for Moldflow it would be nice if there is an option to utilize GPU for the computation. As probably most of the professional user's workstations will have a GPU like Quadro / GeForce equipped it will be great to take advantage of it which may reduce analysis runtime.

Also, I am not sure but just a Wild Idea that it would be better if there is a way to parallelize CPU + GPU core for solving.

Related Articles.

1. https://developer.nvidia.com/blog/computational-fluid-dynamics-revolution-driven-by-gpu-acceleration... 
2. https://www.ansys.com/en-in/blog/unleashing-the-power-of-multiple-gpus-for-cfd-simulations 
3. https://blogs.sw.siemens.com/simcenter/gpu-acceleration-for-cfd-simulation/ 

Regards,

Pavitharan V

[ryan.fuhr]

As I recall, this was an option for a while (circa 2014?), but was removed. I am not sure why. May have been about available RAM?

That said, there are now GPUs (and related cards designed for AI processing) that have large amounts of RAM available. e.g. latest Quadro (RTX 5000/6000) and Fire Pro models have 32 and 48 GB, the Nvidia A100 has models with 40 and 80GB, though I think it's designed to be used in a VM, which as I recall Moldflow does not like.

Almost all of our solves use 32GB of RAM or less.

[bernor_mf]

@ryan.fuhr , @PAVITHARAN_V 

Right, this was available for a while. Some 14 years ago...

Computation by GPU - DEVELOP3D

One reason I believe for disrupting this, was that PC's performance increased. More CPU cores, more RAM.

The benefit of using GPU for solver where maybe then too small.

Regards,

Berndt

[raalteh]

Back story...

As per Berndt's comment above. Moldflow was one of the early GPU technology adopters, but chip designs were optimized (with tech derived from the GPU's). We worked with GPU providers at the time to get more performance out of the system but were not able to do so. We ended up with strange situations that when the GPU's were engaged, the CPU's often needed to wait for the GPU's and effectively slowing the solve down.

At the time Tesla GPU cards could be very expensive and ... give no performance so we made the call to disable the tech.

Recently we did have another look at this potentially found a way to get performance out of GPU cards again. This is still an investigation that would need to go through commercialization but this looked promising.

[raalteh]

The idea is understood and there is some initial evidence is worth looking at again.

[jase1973]

Bump

Any update on this. 

Using the Moldflow Insight on a server to move heat/noise away from User computers and the solver is not leveraging the GPU

[raalteh]

Hi,

A quick update is that we did spend time in the last year and a half to enable the GPU for (3D flow) solves, but we did not release it because the results were not matching our expectations. WE do have two GPU solve initiatives on the docket though; one I have very high confidence in, the other being a bit more speculative.

I agree massively with the sentiment, but it's not just a switch to turn on. 

[jase1973]

Ok, so not implemented as yet. do we have a timeframe?

We have a 32 core server that is going 100% when passed a job. (which runs quicker on a laptop)
Before adding the GPU we researched whether it would help and all the forum posts show how to leverage the GPU with autodesk, little dissapointed it's made no difference.

[raalteh]

I'm sorry we are disappointing you (truely...). I have no timeline, but a completely from the ground up new flow solver may be a requirement (which is a huge development).

Taking a step back to your 32 core comment. The 3D flow solver does not benefit much beyond 8 cores for an analysis. If you want to actively manage your core usage instead of the 'automatic' option I would limit it to 8 to be honest. It would likely help keeping your CPU's from overheating and going into a low performing state and your solve could finish quicker. Just a thought

I completely realize this is an unsatisfactory advice but I make it with the best intentions.

Hanno

[ryan.fuhr]

@jase1973  The testing is a bit dated now (2023 solvers I think), but I shared benchmarking results a while back here: 

https://forums.autodesk.com/t5/moldflow-insight-forum/performance-testing-results-various-cpus-and-p...

That will provide some guidance on # of cores per solve as a function of clock speed, RAM bandwidth, etc.