Hi, I am trying to perform eigendocomposion on ~20000 Hermitian sparse matrices with shapes of up to 1e6. I have looked at the docs of cuSOLVER and cuDSS but the only function that solves eigenproblems is cusolverSp<t>csreigvsi(), which can only find an eigenvalue and eigenvector that converges the initially given ones instead of finding all eigenvalues and eigenvectors.
Is there any other method that can be used other than turning sparse matrices to dense matrices?
Thanks!
At dimensions up to 106, that wouldn’t work, as storage requirements would be in the TB range. I am curious: what package would you (or do you currently) use to perform the desired task on the host, that is, using only the CPU?
Is it a 1e3 × 1e3 or a 1e6 × 1e6 matrix?
They are 1e6x1e6 matrices unfortunatly.
We are normally analyzing a system piece by piece so we didn’t needed to process this big matrices before but we think that analyzing the system as a whole (which can generate matrices up to 1e6x1e6) may produce better results.