Neuromorphic inference circuits using emerging devices (e.g. ReRAM) are very promising for ultra-low power edge computing such as in Internet-of-Thing. While ReRAM synapse is used as an analog device for matrix-vector-multiplications, the neuron activation unit (e.g. ReLU) is generally digital. To further minimize its power and area consumption, fully analog neuromorphic circuits are needed. This requires Design-Technology Co-Optimization (DTCO). In this paper, we use our Software+DTCO framework for fully analog neuromorphic inference circuit optimization using ReRAM as an example. The interaction between software machine learning, ReRAM, current comparator, and ReLU are studied. It is found that the neuromorphic circuit is very robust to the variation of ReLU, which confirms the importance of DTCO simulation.