-
Notifications
You must be signed in to change notification settings - Fork 0
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
GPU usage slower than CPU #53
Comments
Using profiling, it seems that all parts of the forward/backward loop are slower on GPU. This was for a batch of 10 muons, since profiling is really slow to run (typically we run 100 per batch). I would guess that for our kinds of operations, we aren't able to really benefit from GPU, due to each operation being relatively light (compared to multiplying large matrices in a DNN), and the muon batch sizes a small.
|
PR #52 fixes code to run on GPU, however runtimes are about twice as slow than on CPU.
Possible causes:
Need to examine code. The new PyTorch profiler might help highlight the slow parts of the code.
The text was updated successfully, but these errors were encountered: