[tyler1008]

My specific area of usage is for Fluent, heat transfer simulations.
most of our labs work are small sized.
Currently, we have a cluster, and each node has epyc 7281, 2*16core(64thread) computing node with 256 gb of ram.(8*2=16channel)

So here is the deal, Our lab's work mesh sizes are not that big, which can be ran from office PC with i7 +8*2=16gb ram. of course takes some more time.

Now, what I am thinking is, changing the PSU of office PC, and installing used Quadro GV100, which is $300~600 for used one from ebay.
https://www.ansys.com/content/dam/am...2023-r2-ks.pdf
From this source, page 16, it clearly states that one tesla v100 is equvalent to 150-220 core, which will be 3-4 times faster then current 1 cluster node(64core).

With my checklist,

Does your CFD package support GPU acceleration?
Yes
Do the solvers you intend to run benefit from GPU acceleration?
Yes
Are your models small enough to fit into GPU memory? Or vice versa, how much VRAM would you need to run your models?
Yes. mesh below 10M, 16gb
Does GPU acceleration for your code work via CUDA (Nvidia only) or OpenCL (both AMD and Nvidia)
Not sure

Which GPUs are allowed for GPU acceleration? Some commercial software comes with whitelists for supported GPUs for acceleration, and refuses to work with other GPUs.
Aiming to buy used V100 16gb.

Single or double precision? All GPUs have tons of single precision floating point performance. But especially for Nvidia, only a few GPUs at the very top end also have noteworthy double precision floating point performance.
V100 16gb has 17tflops single,
8.3 tflops double precision(1:2 Fp64)

Therefore, I think this is valid choice.
how do you guys think? Please point out the points that I am missing.

+ edit

Now I am thinking of building with 2 V100, since 3 HPC license is upt to 168 SM, and one V100 equals 80SM.
Funnily, 1HPC=upto 48SM, 2HPC =up to 72SM, 3HPC=up to 168SM, so there is no change between one and two V100 license.


++edit 2

The conclusion how I had this was,
23r2 and later fluent is very supportive with multiple GPU.
Seems from 2m to 250m mesh cells, adding GPU is linear to performance, <- make sense cause adding badnwidth and performance at once

[JBeilke]

I wanted to try it out myself, but the models that plagued me with long calculation times were too big for the card's RAM. However, I would appreciate it if you tried it out and then let us know how it worked :-)

[flotus1]

Let's put it this way: if a Quadro GV100 doesn't get decent results, no other card will. It is pretty much made for this use-case.

[JBeilke]

What about the compatibility of this "old" card and the current versions of CUDA?

[tyler1008]

Well, The ppt slides are from the ansys, which was rolled out based on ansys 23r2, which was fully supported for GPGPU.
I guess the compatibility will be fine, as V100 was with one of the first tensor cores which also boosted FP64.