Abaqus gpu acceleration

[FluentFriend]

Hello!

recently i've found out about the option in abaqus to use the gpu for acceleration (a dream of mine becomes true )

at the moment however our company is only using weak quadro cards resulting in even longer simulationtimes using the gpu.

there are several announcements about the gpu acceleration (2.1 faster with 5M DOF)but hardly any further information.

my question is:
has anyone allready some experience with abaqus and gpu acceleration? (especially with firepro cards because all the quadros i expect are much weaker)
how important ist the gpu mem?
how does it scale with adding more gpus? (crossfire)
is it also faster with much bigger models? (some of our jobs need about 250GB RAM)

looking forward to hear from you.

 

[csk62]

GPUS are slower for smaller jobs because it takes longer to transfer the memory. I would only use them for larger jobs. The benefit of GPUS, in my opinion is that it only takes one token per use, whereas the CPU-token tradeoff can be much more expensive.

 

[FluentFriend]

thanks for your advice!

what is for you a larger job?
most of our models have a size of 7-10M DOF, is that already big enough for gpu acceleration?

 

[csk62]

DOF isn't necessarily the only indicator of runtime. The complexity of your contact definitions, material definitions, nonlinearities can cause the analysis to move at snail's pace as well.
I seldom use GPUs because ours are so crappy I hardly think it makes a difference (If not worse, which I think you have observed as well). I think I remember someone telling me Abaqus is re-configuring the licensing/token system for GPUs, and I think ANSYS has already done so.

I am still waiting for the day that we can bring some new computing so I can really test out the GPU accelerator as well. But Im afraid I dont have too much experience with running GPUs for Abaqus jobs.

 

[FluentFriend]

That's true, DOF is not representing the whole model but at least it's a hint how big the model is

The licensing is already changed, one GPU uses one normal Token, no need anymore for special GPU tokens.

Thanks for your input!

It's really strange, GPU acceleration has such a huge potential but i can't find anybody who really uses it every day.
All the infos i've found are from NVIDIA and AMD and that makes me douptfully...

 

[loki3000]

the cheapest option is nvidia titan at $1000.
note that it doesn't have ecc though, but if you run it on a workstation, it will be ok anyway.

 

[FluentFriend]

thanks for your advice, however a titan ist much too weak. (only 1.3 TF double precision)

compared to the amd fire pros, all the nividas are much weaker (that's why nvidia made cuda to outperform amd through easyer api)

i'm not looking for a cheap solution, we whant performance.

the main question is, how much ist the improvement with a new amd firepro.

2 times faster in nearly all simulations for 3000 - 6000$ and we'll buy it