The requirement of higher data rates motivates the adoption of high-order modulations and pulse shaping in optical systems. This paper proposes an efficient architecture for FPGA-based parallel fractional sample rate conversion that has an optimum computational cost. The architecture supports non-integer oversampling factors, which allow increased flexibility when compared to schemes limited to adopting one or two samples per symbol, for example. The presented example is evaluated for a spectrally-efficient 400G optical supper-channel (three sub-channels with 16-QAM modulation) with different pulse shaping filter implementations using an FPGA working with a clock of 156.25 MHz. The results also illustrate the impact of the pulse shaping when a non-integer oversampling is adopted.