QunaSys Tech Blog

Quantum-Selected Configuration Interaction (QSCI) uses a quantum device to propose important electronic configurations, or determinants, and a classical solver to build and diagonalize the Hamiltonian in that subspace. Larger selected subspaces can improve the resulting energy, but only if the classical solver can handle their Hamiltonians. At QunaSys, we distributed both classical stages in a QSCI implementation being developed for QURI SDK Enterprise, with each process holding only part of the matrix. This removed the single-process memory ceiling and enabled runs with four million determinants on Fugaku. On H18 at one million determinants, Hamiltonian construction scaled 11.8x from 4 to 96 processes. On four cytochrome P450 models, our solver ran 8x to 17x faster than the tested Selected Basis Diagonalization implementation at the largest subspace sizes compared, with energies agreeing to better than 1 mHa. At the largest subspace measured for each system, diagonalization dominated the classical runtime, while Hamiltonian density predicted cost better than qubit count. Pushing to eight million determinants showed that peak memory use during assembly is the next bottleneck. Our next target is a matrix-free eigensolver that never forms or stores the sparse matrix.