You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have some questions regarding training runtime and the number of GPUs.
In your paper on table 4 (page 22), it shows the actual training runtime of each experiment.
Are these reported runtimes are measure by running w/ multiple GPUs?
Especially for Navier-Stokes equation, you also experimented w/ 8 GPUs, and I want to know whether the 68.29 it/sec of runtime was achieved by multiple GPUs.
Thank you.
The text was updated successfully, but these errors were encountered:
Hello, thank you for sharing your wonderful work!
I have some questions regarding training runtime and the number of GPUs.
In your paper on table 4 (page 22), it shows the actual training runtime of each experiment.
Are these reported runtimes are measure by running w/ multiple GPUs?
Especially for Navier-Stokes equation, you also experimented w/ 8 GPUs, and I want to know whether the 68.29 it/sec of runtime was achieved by multiple GPUs.
Thank you.
The text was updated successfully, but these errors were encountered: