A 0.055pJ/bit/dB 42Gb/s PAM-4 Wireline Transceiver with Consecutive Symbol to Center (CSC) Encoding and Classification for 26dB Loss in 16nm FinFET

The growth of data-intensive applications such as large language AI models has increased the demand for higher data throughput in wireline links. Due to the bandwidth-limited nature of the wireline channel, increasing the data rates across the same physical distance of the communication channel results in more inter-symbol interference (ISI). Consequently, more channel equalization is required to compensate for ISI, which increases the energy/bit of the communication link. Researchers have discovered that machine learning (ML) inspired approaches [1]–[2] including feature extraction and classification provide a more efficient solution for compensating the channel loss compared to the conventional equalization techniques like FFE, DFE, and CTLEs [3]–[7]. However, the prior works on ML inspired links are limited to NRZ modulation only. In this work, we introduce an energy-efficient ML inspired transceiver that leverages feature extraction and classification to transmit encoded PAM-4 data across a wide range of channel loss (13dB to 26dB) while maintaining BER < 10⁻¹¹ without using any conventional equalizers. Additionally, we propose a data encoding scheme, consecutive symbol to center encoding (CSC) to encode PAM-4 and provide identifiable attributes to the transmitted signal, which helps to increase the channel loss compensation range and reduce the complexity of the classifier. Since ISI is a deterministic non-ideality, the proposed decision-tree based classifier (on-chip) is designed to learn both the channel characteristics and the CSC data encoding, enabling it to accurately detect the original transmitted data in the presence of ISI with a latency of only 10 unit interval (UI). The decision tree classifier operates with a low-power feed-forward architecture without any feedback timing constraints, allowing the proposed transceiver to achieve 0.055pJ/bit/dB, which is ~2x lower than prior work [3]–[5] while compensating for nearly the same channel loss.