Approximate Computation using Neuralized FPU
Neural networks can be used as function approximators to improve the energy efﬁciency, performance, and fault-tolerance of traditional computer architectures. To maximize these improvements the granularity of the function must be as large as possible. This work-inprogress abstract explores the lower limits of neural network function approximation by replacing individual ﬂoating point multiplications with multilayer perceptron neural networks. We show that this ﬁne-grained approximation technique provides 2%-100% application output accuracy for multiple applications in the PARSEC benchmark suite across varying network topologies.