We investigate unreliable computation problem under cost and fidelity constraints. We introduce a framework, where we observe a noise-corrupted version of correct computation while incurring an associated cost. We consider several cost models formalizing the relationship between fidelity of an unreliable computation and its cost. We demonstrate that a single unreliable computation incurs less cost than a fusion of several less costly and less reliable computations while achieving same MSE performance under concave and linear costs. We show that fusing several less costly and less reliable computation yield a better cost-performance tradeoff than a single unreliable computation under convex costs.