A function-based memory partitioning (FBMP) method is proposed and adopted in a low power digital signal processor (DSP) for cochlear implants (CIs). Based on preanalysis of intended application programs, instructions and data are divided into functional groups with different accessed frequencies. Monolithic instruction and data memories are then partitioned into multiple sub-banks and assigned to functional groups according to their accessed frequencies. The FBMP method can reduce both power consumption and silicon area overhead. The implementation result on the CI-dedicated DSP shows that this method can save about 50% of the power at best with little additional area cost.