As a part of other work I need to solve relatively large (~1E5x1E5) and sparse (~100 non-zero elements in each raw in few blocks) hermitian eigensystems. Usually only few eigenvalues+vectors are needed, but with high precision. Currently I am using Arpack (Arnoldi method with shift-inverse when precision is preferable or spectrum folding when size is important). As an option plan to use TRLan (thick restart Lanczos) and try Chebyshev filtering instead of spectrum folding.
Probably, newer methods exist for this purpose? I am not an expert in CS, maybe somebody has some clues on recent progress in this field?