Mixed-signal neural network branch prediction,” unpublished manuscript

Accurate branch prediction is essential for modern microprocessors in order to maintain high in-struction throughput. Neural networks have shown great promise in branch prediction. However, digital implementations require complex circuitry leading to low cycle times, which have precluded the use of these predictors in actual microprocessor designs. This paper evaluates the potential for leveraging analog circuitry to efficiently implement neural network branch predictors. In particular, we propose to implement a perceptron branch predictor using simple analog blocks including a resistor network for the synaptic multiply/add function and a comparator for the non-linear output function. We have simulated this circuit using a C model embedded into SimpleScalar that calculates circuit voltages and currents. We have found that the prediction accuracy of the analog branch predictor is highly tolerant to resistor mismatch and noise. In the presence of realistic resistor matching and noise conditions, it achieves virtually the same accuracy as a digital implementation. The simplicity of the analog perception should enable it to operate at several GHz with lower power consumption than a digital implementation.