Page 1 of 1

Should I select Double Precision or Single Precision GPU card for VASP GPU Version?

Posted: Thu May 02, 2024 1:48 pm
by hszhao.cn@gmail.com
Dear VASP Community,

I am considering the GPU-accelerated version of VASP. I understand that accuracy is paramount in these calculations, and I am trying to decide between a double precision and a single precision GPU.

Could you please provide advice on whether a double precision GPU is necessary for reliable results with VASP? Are there particular simulations within VASP that absolutely require double precision GPU card? I would greatly appreciate any insights or shared experiences from the community.

Thank you for your guidance and support.

Best regards,
Zhao

Re: Should I select Double Precision or Single Precision GPU card for VASP GPU Version?

Posted: Thu May 02, 2024 3:18 pm
by michael_wolloch
Dear Zhao,

VASP relies very heavily on double-precision operations. Therefore GPUs with good FP64 performance are necessary to obtain decent speeds with VASP, as is stated in the wiki page on the OpenACC GPU port of VASP.

Thus "Gaming cards" are not useful, since they limit FP64 performance to 64 times less than FP32 performance usually. This might also be true for some "Professional" cards, like the RTX 6000 Ada generation. But you have to be careful even for "Data Center" GPUs: The A40 has roughly 4 times better FP32 performance than the A30 for example, but ~10 times worse FP64 performance. This is because the GA102 chip used in the A40 only has 2 FP64 units per streaming multiprocessor, in contrast to the A30, which is GA100-based, the chip that is also used in the flagship A100 card.

Please let me know if this completely answers your question, so I can lock the thread,
Michael

Re: Should I select Double Precision or Single Precision GPU card for VASP GPU Version?

Posted: Fri May 03, 2024 7:03 am
by hszhao.cn@gmail.com
Dear Michael,

Thank you very much for your valuable comments and explanations. For now, I have decided to select the GPU cards as follows:

1. For a single GPU card used on my workstation, I will use the GP100 16G version.
2. For GPU card array-based GPU servers, I will use the v100 sxm2 16/32 G version to do the trick.

See here for the related discussion.

Regards,
Zhao