-
Notifications
You must be signed in to change notification settings - Fork 11
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
Different results on different GPUs #28
Comments
Do you know if this is specific to interpax? It's likely it's a more general JAX issue (or really a CUDA/XLA issue) that things get compiled differently for different hardware, see jax-ml/jax#20371 and jax-ml/jax#10674 (comment) Also, is the error uniformly bad for all points being interpolated, or is it localized in some way? |
Hi! Thank you for the reply. I believe it's related to a general JAX-related issue since I could not observe machine-specific implementation in interpax. Here I attached two images, which present relative pointwise error This is for 4090 with single precision, and this is for 4090 with double precision. For the left vertical edge of the figures, In fact, my query points |
Can you share some code/data that seems to reproduce the issue? I don't have access to either of those GPUs but I can try some others and see if its a more general issue. |
Hi, I think I found the cause. I ran the test with It might be related to this issue (default TF32 overriding of JAX): |
Hi f0uriest,
I encountered an issue that interpolation results vary along different machines.
I used a 1d interpolator with the monotonic method, allowing
extrap=True
.test machines: [CPU, RTX Titan, RTX 4090].
reference machine: CPU with double precision (x64).
below table presents relative$L^1$ error:
abs(a - b).sum() / abs(b).sum()
Since I used the same
(xq, xp, yp)
, the errors of each row must coincide, respectively.However, as you can see, interpolation on RTX 4090 with single precision produced quite an inaccurate result.
Do you have any ideas on this?
The text was updated successfully, but these errors were encountered: