Although baud-rate clock and data recovery (CDR) such as Mueller-Müller (MM) CDR is adopted to ADC-based receivers (RXs), it suffers from false-lock points when the RXs handle PAM4 data pattern because of the absence of edge data. In this paper, a false-lock-aware locking scheme is proposed to address this issue. After the false-lock-aware locking scheme, a clock phase is adjusted to achieve maximum eye height by using a post-1-tap parameter for an FFE in the CDR loop. The proposed techniques are implemented in a 56-Gb/s PAM4 transceiver. A PLL uses an area-efficient “glasses-shaped” inductor. The RX comprises an AFE, a 28-GS/s 7-bit time-interleaved SAR ADC, and a DSP with a 31-tap FFE and a 1-tap DFE. A TX is based on a 7-bit DAC with a 4-tap FFE. The transceiver is fabricated in 16-nm CMOS FinFET technology, and achieves a BER of less than 1e-7 with a 30-dB loss channel. The measurement results show that the MM CDR escapes from false-lock points, and converges to near the optimum point for large eye height.

Many studies of deep neural networks have reported inference accelerators for improved energy efficiency. We propose methods for further improving energy efficiency while maintaining recognition accuracy, which were developed by the co-design of a filter-by-filter quantization scheme with variable bit precision and a hardware architecture that fully supports it. Filter-wise quantization reduces the average bit precision of weights, so execution times and energy consumption for inference are reduced in proportion to the total number of computations multiplied by the average bit precision of weights. The hardware utilization is also improved by a bit-parallel architecture suitable for granularly quantized bit precision of weights. We implement the proposed architecture on an FPGA and demonstrate that the execution cycles are reduced to 1/5.3 for ResNet-50 on ImageNet in comparison with a conventional method, while maintaining recognition accuracy.