Do you know if there is any app_config options to pass a command line option to disable the throttling?
there is none--nor could they be, as the application is not requesting the throttling.
Quote:
I was not aware that EAH messed around with the frequency.
It does not.
What you are seeing in the frequencies (both GPU and clock) is that NVidia (not Einstein's) control overhead lowers the clock rates by a lot when there is nothing for the GPU to do and has not been for a little while.
In the Turing cases, what appears to happen very early on is something which deflects progress from the normal path, so that all measures of GPU activity, including clock rates, and GPU load, show that instead of the normal startup of steady processing, there is less than three elapsed seconds (perhaps much less) of GPU engagement during the start sequence.
One of the Turing puzzle pieces is that for Windows cases it takes more than ten additional elapsed seconds before these tasks get an error termination. And in the Linux cases reported so far, it seemingly never terminates at all (and in the Linux cases the GPU seems active, though not accomplishing any progress as detectable by checkpoints.
Please reread my response at
)
Please reread my response at the beginning of this thread. EAH doe not affect the frequency.
Well, if you say that it
)
Well, if you say that it doesn't affect the frequency, then no need to reread your response. Thanks
rjs5 wrote:Do you know if
)
there is none--nor could they be, as the application is not requesting the throttling.
It does not.
What you are seeing in the frequencies (both GPU and clock) is that NVidia (not Einstein's) control overhead lowers the clock rates by a lot when there is nothing for the GPU to do and has not been for a little while.
In the Turing cases, what appears to happen very early on is something which deflects progress from the normal path, so that all measures of GPU activity, including clock rates, and GPU load, show that instead of the normal startup of steady processing, there is less than three elapsed seconds (perhaps much less) of GPU engagement during the start sequence.
One of the Turing puzzle pieces is that for Windows cases it takes more than ten additional elapsed seconds before these tasks get an error termination. And in the Linux cases reported so far, it seemingly never terminates at all (and in the Linux cases the GPU seems active, though not accomplishing any progress as detectable by checkpoints.