How to perform eigenvalue decompositions for large matrices with Python

[Curefab]

There also was once (10.1) a bug, that a quadratic buffer size requirement was returned, but should have been fixed by now (11.0).

CuSolver: can't compute SVD on tall matrices with Cuda 10.1, bufferSize grows quadratically

Hi all, I recently upgraded to Cuda 10.1, and I can’t compute general SVD of tall matrices anymore, that is cusolverDnSgesvd with large m, tiny n. After some inspection, I noticed that cusolverDnSgesvd_bufferSize is quadratic, but I believe the memory needed to compute SVD should be linear. Here is the plot of values in log scale: sgesvd_bufferSize - Imgur I actually get an overflow for m larger than ~32k, which is not that tall. Is there any known workaround?

cusolverDnDgesvd_bufferSize - hugh increase in buffer size requirements with 10.1

With CUDA ToolKit 8.0 it is possible to perform SVD on much larger matrices using “cusolverDnDgesvd” than with CUDA ToolKit 10.1. Testing with a Telsa P40, the results as per the code below indicate that for a 40000 X 5000 matrix, the working space required is: 14.833 GB for 10.1 0.200 GB for 8.0 Sadly, the 10.1 version of “cusolverDnDgesvd” runs something like 20X faster than the 8.0 version so it would be great to use the 10.1 version, but I need to be able to handle matrices that are lar…

sgesvd_bufferSize int32 overflow with CUDA 10.1 · Issue #2351 · cupy/cupy · GitHub (that issue describes both problems)