Advanced search

Message boards : Graphics cards (GPUs) : Application GPU precision

Post to thread Subscribe Sort
Author Message
TRINITAS
Send message
Joined: 20 Apr 12
Posts: 3
Credit: 13,845,248
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwat
Message 61083 - Posted: 25 Jan 2024 | 19:21:15 UTC

Hi all,

I would like what is the calculation precision used by GPUGrid applications? It won't be FP64 by chance?

Thanks :)

ID: 61083 | Rating: 0 | rate: Rate + / Rate - Report as offensive    Reply Quote
Ian&Steve C.
Avatar
Send message
Joined: 21 Feb 20
Posts: 1079
Credit: 40,231,533,983
RAC: 11
Level
Trp
Scientific publications
wat
Message 61086 - Posted: 25 Jan 2024 | 19:48:24 UTC - in response to Message 61083.

ACEMD3 and ATMbeta are primarily FP32 as far as I'm aware, i think ATMbeta also uses some small fraction of FP64 since in it OpenMM is configured for mixed precision and the OpenMM docs say:

f it is set to “mixed”, forces are computed in single precision but integration is done in double precision.

but i don't really know how much of the overall calculation is that integration part. Since strong DP cards aren't miles faster here, I assume its a relatively small portion that's done in FP64

The Quantum Chemistry on GPUs project however does use a large portion of FP64, and strong FP64 GPUs are much faster for that. they also use Tensor cores (if available) as far as I know.
____________

ID: 61086 | Rating: 0 | rate: Rate + / Rate - Report as offensive    Reply Quote
TRINITAS
Send message
Joined: 20 Apr 12
Posts: 3
Credit: 13,845,248
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwat
Message 61088 - Posted: 25 Jan 2024 | 20:30:59 UTC - in response to Message 61086.

Ok thanks you :)

So there is mixed precision FP32 and FP64

ID: 61088 | Rating: 0 | rate: Rate + / Rate - Report as offensive    Reply Quote
Post to thread

Message boards : Graphics cards (GPUs) : Application GPU precision

//