Equalization with Neural Network Circuitry for High-Speed Signal Link

Abstract

Traditional data link relies on an open eye diagram to ensure correct data transfer. There are many factors, including jitter, noise, crosstalk, channel bandwidth/filtering, etc. that can shrink the eye diagram. Equalization (EQ) including the finite impulse response (FIR) EQ, continuous time linear EQ (CTLE), and decision feedback EQ (DFE) are used to improve the eye diagram. However, when the aforementioned factors are so severe that the eye is closed, data transfer will fail. We propose a new approach that does not depend on an open eye diagram. The neural network (NN) circuitry, which we call an NN interpreter, is used to replace the EQ mechanism. This new solution can ensure correct data transfer even when the eye diagram is closed with the traditional EQ mechanism. The NN interpreter takes the voltage (or current) waveform in length of several unit intervals (UIs) and make the decision of 0 or 1 based on the entire waveform instead of just comparing the voltage in the cursor UI with the reference voltage. Simulation results with a DDR5 channel show over 30% improvement in data rate with the proposed method, that is, the data rate of 6 GT/s with traditional equalization method is improved to 8 GT/s with the NN approach.