Learning Forum

[info]

Hi,

I wanted to know if we could have more information on the solver used by Discovery Live as there is no official doc yet? Some sources claim that the solver is based on the lattice Boltzmann method, can you confirm ?

 

Concerning the GPU fidelity: ANSYS affirms that DL works well with Quadro and Geforce. Considering the processing power of each one in single and double precision and that the solver seems to run as fast on one card as on the other, one can assume that the solver is running single-precision (FP32) ? Or does the solver adjust itself depending on the card: 

 

Geforce => SP (FP32)

Quadro => DP (FP64)

 

Can you comment on this ?

Thank you

[Naresh Patre]

Hello Quentin Lux Regarding your first question on solver technology, please check out below post: What simulation technology does Discovery Live use?   Regarding your second question on GPU, I will check with team and get back to you.

[Naresh Patre]

Hello Quentin Lux

Regarding your second question on GPU, Discovery Live is designed to perform well on a wide range of NVIDIA GPUs including both Quadro and GeForce cards.

[info]

Hi Naresh Patre 

Thank you for your answer, that was my understanding too but that is not really my question: considering the fact that Geforce perform badly in double precision (FP64), is the solver running single precision (FP32 / 32 bits ) ?

[Naresh Patre]

Quentin Lux There is no other information to share apart from the one posted in previous reply. If you are noticing any specific issues while running Discovery Live, please share the details and we will look into it. 

[info]

Hi Naresh Patre ,

It is not really an issue but a matter of confidence in the results and it is important to know, it really should not be a secret.

Apart from this, it could also help decide on the type of card to purchase since some really expensive Tesla card have poorer single-precision flops than gamer Geforce FX: for example a 1080Ti rated at 10000 Gflops at 1k$ vs a Tesla P100 at only 8000 Gflops for 6k$ ?

According to the literature, most of GPU computing on LBM is performed single-precision so I will go ahead and assume that it is.

[brajagopalan]

Quentin Lux Please go to this Ansys post for the single precision subject. As for the GPUs, Quadro GPUs are recommended for physical desktops. Go to this page to understand why Quadro GPUs.