[Daveo643]

Hello, I hope someone here is able to help give me a definitive answer to my queries. Fluent 14.5 has support for utilizing supported NVIDIA GPUs to off-load some computation tasks from the CPU with an HPC License (which my institution has). What is not clear from the literature I've read is whether this capability is for a specific set of NVIDIA cards per ANSYS or more general as suggested by NVIDIA. The ANSYS sales reps with whom I'm made this inquiry are also unaware or unsure of this. The published demonstrations I've see utilize Tesla K20 or C2050, e.g. see the below links.
http://www.ansys.com/staticassets/AN...ry.pdf#page=19
http://www.ansys.com/staticassets/AN...ry.pdf#page=20

I know there's always a distinction between the definitions of "qualified," "supported," "certified," "tested" and variants when talking about engineering software. We have workstations that variously have NVIDIA Quadro cards and I'm doing my Fluent CFD work on a high-end laptop with a Core i7-2760QM CPU, Nvidia GeForce GTX580M with 2GB video RAM and 32GB system RAM, but not one of the cards listed in the above references. For my purposes, I don't need ISV certification but any speed-up will help (running up to a 5.5 million cell transient, turbulent flow model that takes upwards of 48 hours to run on my local machine).

My questions are:
1.) Can I exploit GPU acceleration on the abovementioned GPU not expressly listed by ANSYS?
2.) Does the software poll the GPU ID to determine whether to support GPU acceleration?
3.) Can GPU acceleration be forced to be enabled?

Thanks very much in advance.

[Daveo643]

PS: Simulations are expressly double precision if that changes anything.

[CDollarsign]

I am also in need of this info, can't find much information on this. How do I know if my GPU is being used, I have a Tesla 2045...

[RodriguezFatz]

As far as I know the GPU can just be used for some special model (radiation?) but not for the NS-solver. Are you aware of that? Thus, every common simulation does not improve by CUDA.

[Bollonga]

I'm also very interested in GPU-acceleration.

Originally Posted by RodriguezFatz

As far as I know the GPU can just be used for some special model (radiation?) but not for the NS-solver. Are you aware of that? Thus, every common simulation does not improve by CUDA.

I've found this, maybe it's useful:
http://developer.download.nvidia.com...luent_SC12.pdf
In page 20 it says Ansys 14.5 solves radiaton heat and AMG solver just beta version.

http://www.microway.com/pdfs/NVIDIA-...r%20Facing.pdf
Here it's similar, for Ansys Fluent just radiation heat transfer.

I guess this is quite limited!

[CDollarsign]

What a disappointment. My lab computer has a Tesla 2045 just waiting to crunch some numbers...

[sadegh1068]

In ANSYS 15 you can use GPU for solving NS-equation.
you can see this text in ANSYS web site:

"Engineers always need faster solutions, and ANSYS investigates all technologies that will help them do so. At release 15.0, ANSYS Fluent supports solver computation on GPU. This can lead to a speedup of up to 2.5 times.GPU support for the 3-AMG coupled pressure-based solver demonstrates the ANSYS commitment to allowing customers to leverage new and evolving technology, such as GPU, for faster simulation."

[Zaktatir]

You can use GPU but if you get problems on not certified cards you won't have a warranty to get support upon this. This is fortunately quite limited that means the support will afford you to switch GPU if you have problems let's say in speed up or an communication issue.

GPU is not support for multiphase flow.

[davidiitk]

Hi,
I am working on FSI problem. I don't have a graphics card except inbuilt intel. please recommend an appropriate one, mainly for post analysis in fluent.
my range is up to Rs. 50,000/-

[dr.chris]

We have just been doing a little research on this issue.

The word from Ansys is that due to memory limitations on GPU chips they are not suitable for the efficient partitioning of meshes needed to make CFD parallelisation practical. So currently they are limited to ray tracing and the multi-grid solver for the coupled pressure-velocity formulation.

The technology only works with a limited number of GPU cards and its not cheap either. GPU's are more useful for FE calculations but as far as CFD goes it may well be one for the future but not really for now.

[TomP]

Does anybody has some experience with accelerating Ansys Fluent 15 with a Geforce GTX Titan? I have Tesla cards in my cluster but I would like to make a case for an extra card in my workstation for test runs. I just need to know if it works, I'm not interested in the fact that the card is officially supported by Ansys.

Tom

[robmuggleton]

I can confirm the my nvidia k3100m works with the fluent solver, using coupled solver. Watching gpu with msi afterburner, gpu fluctuates between 0 and 80%, not at constant percentage.

[orangesky]

The document's been updated from ANSYS that with more 'supported' GPU's and a number of other 'tested' GPU's for FLUENT

http://www.ansys.com/staticassets/AN...ed-summary.pdf

I'm ordering a K2000 card so I'm wondering if anyone has yet used the GPU processing on R15? If so, what was it used for and how were the results?

[jmorales]

Originally Posted by TomP

Does anybody has some experience with accelerating Ansys Fluent 15 with a Geforce GTX Titan? I have Tesla cards in my cluster but I would like to make a case for an extra card in my workstation for test runs. I just need to know if it works, I'm not interested in the fact that the card is officially supported by Ansys.

Tom

I have doing a lot of tests using Nvidia GTX Titan, it has more performance as Tesla K20 with lower adquisition costs, the problem here is that Titan is not passive GPU as Teslas or Quadro, so they can be damaged by the continuos load of simulations, my research conducts to an efficiente thermal distributions of air flow and liquid cooling along the Cases to improve the performance and the reliabilty of the Titan GPU with extensive load journals.
Go for Titans if you dont want expend thousands of dollar on Teslas GPU, at least you really need it.

[Daveo643]

Following up if anyone has VERIFIED that Fluent and only Fluent works with the GTX Titan. Screenshots or results from the Fluent command line that shows the card being exploited would be much appreciated. I tried contacting jmorales above but no response.

I'm now on V15.0.7 FWIW.

[valahian1]

In my case (Quadro 6000 and Tesla C2075), just Quadro works with FLUENT. How do I now:
- in Fluent Launcher, set 2 GPGPUs (parallel processing);
-launch Fluent; if there are some problems, you will see the message:
"Inhomogeneous process distribution on multiple machines.
Or processes per machine not an exact multiple of GPGPU's per machine.
Or not enough GPGPU's per machine.
GPGPU computing disabled."
The GPU's computing is disable and you will run your case just in parallel, with yours CPUs.
- if you don't receive the message above, everything is OK, your simulation will run on GPUs too.
- during solver process, open Nvidia Control Panel/Manage GPU Utilization and see which GPUs are working.

Do not expect to a high improvement using GPU. If you install desktop gadgets like GPU Meter and a CPU Meter to see working cores, you will observe how poor is the contribution of GPU in computational process.
Just Quadro works for me (12 processes and 1 GPGPUs per machine) and Tesla lazes there making me frustrated.

Have good and fast simulations!

[Daveo643]

The reason why Fluent preferentially runs on your K6000 instead of the C2075 might be because each Nvidia CUDA-enabled GPU has a different Compute Capability level "CC". The K6000 is at CC 3.5 while the C2075 is 2.0 according to here:
https://developer.nvidia.com/cuda-gpus

Fluent probably looks for the highest CC GPU installed if there's more than one and goes preferentially for that. That's my hypothesis from reading around, particularly after stumbling on this:
http://www.semiaccurate.com/forums/a...hp/t-7808.html

In your case, you're probably better off computing off the K6000 than your Tesla card since the former is more recent and powerful.

There's no question that any GPU on ANSYS's supported list will work. My question concerns those GPUs/cards not explicitly on the list. Not surprisingly, neither ANSYS nor Nvidia when I contacted them directly were willing to answer this question directly , instead insultingly referring me to the aforementioned list as if I haven't already done that -- several times since I first brought up the question in March 2013 -- or insinuating that I can't read.

I'm looking at Tesla K80 (not on the list) or the GTX Titan Z. Naturally I'd go for the latter, even getting 2X for about the same cost as one K80. I don't need ECC memory for my purposes, just speeding up a transient internal combustion engine simulation (millions of cells, DP, rke, partially premixed combustion, fuel injection with supersonic flow across the injector throat) over 1000+ timesteps and max 1000 iterations/timesteps to converge). These simulations could take a week to solve so any speed-up, no matter how little, would be useful.

[jmorales]

Originally Posted by Daveo643

Following up if anyone has VERIFIED that Fluent and only Fluent works with the GTX Titan. Screenshots or results from the Fluent command line that shows the card being exploited would be much appreciated. I tried contacting jmorales above but no response.

I'm now on V15.0.7 FWIW.

Hi, sorry for the late, the GTX titan actaully works very good with Fluent 15.0, my bench test has this specs:

Intel Core Xeon 2650L 10 cores/20 threads 1.8 ghz
64gb RAM corsair vengeances non ecc
GTX Titan 2x (not in SLI for GPGPU computing)
Corsair AX 1200W PSU
Liquid Cooler CPU
256gb SSD
2tb HDD

I made test too with the Tesla K20 and the Titan perform much better.

[JKWJD]

to jmorales:

Hi, could you please make some screenshot of parameter of your GTX Titan? If it´s latest generation of "Z" "X" or "Black" which works? I´ve read that current generation of Titan X has worst performance in double precision than previous.

Thak you

[jmorales]

Use the titan normal, black or Z... and Yes, the "Titan X" has lost their Dual Precision Computation capabilities, so, honestly this is not a "titan" because their was made to help in workstation escenerios instead of Quadro or Tesla Processors. Go for a Pair of GTX titan Black for the same 1,000 us dls of the Titan X....

Sorry for the late answer, im in Germany now and i was not checking this forum for a long Time, I still making Research in GPU computing, so if i can help you in other thins, please let me know