This article introduces a wireline receiver (RX) data-path employing discrete multi-tone (DMT) modulation for communicating over electrical links. The DMT RX incorporates a fully digital equalization data-path, with a synthesized and automatically placed and routed digital signal processor (DSP) following a 10-bit time-interleaved pipelined successive-approximation register analog-to-digital converter (TI-PISAR ADC). The prototype RX chip implemented in a 14-nm FinFET process demonstrates a lane data rate of 56 Gb/s dissipating 161 mW including the ADC and the DSP power. The energy efficiency of 1.2 pJ/b for the DSP and 2.9 pJ/b for the entire RX was achieved with the data-rate of 56 Gb/s for communicating over channels exhibiting up to 28-dB loss at 14 GHz with a bit-error-rate (BER) better than 2e-4.
29 August 2018 in IEEE Journal of Solid-State Circuits
A single-channel 12-bit SAR ADC achieving 250–340 MS/s and consuming 4.8–8.0 mW from 0.75 to 0.9 V is presented. At 300 MS/s, the ADC exhibits 61.6-dB peak SNDR and reaches 60.5-dB SNDR and 78.7-dB SFDR with 0.8- $\text{V}_{\text {pp, diff}}$ input amplitude at Nyquist. It consumes 7.0 mW from a single 0.85-V supply, where 3.7 mW is contributed by the reference buffer. The key element is a comparator with an inverter-based preamplifier to achieve low-noise performance with below 1- $\text{V}_{\text {pp, diff}}$ input amplitude. The common-mode (CM) sensitivity of the inverter is counteracted by an SAR-based CM regulation (CMREG). The regulation adjusts the sampled CM to the optimal CM for the maximum inverter gain using a second capacitive DAC. It adjusts the CM on a sample-by-sample basis and, thus, can correct time-varying CM. The implemented CMREG maintains SNDR above 60 dB for a differential input amplitude mismatch of up to 2 dB or a phase mismatch of up to 15°.
19 March 2018 in IEEE Transactions on Circuits and Systems I: Regular Papers
The implementation of a 25.6-Gb/s four-level pulse-amplitude-modulation (4-PAM) reduced-state sliding-block Viterbi detector (VD) is presented. The power consumption of the VD is 105,mW at a supply voltage of 0.7 V, corresponding to an energy efficiency of 4.1 pJ/b. A data rate of 30.4 Gb/s is achieved with an energy efficiency of 5.3 pJ/b at a supply voltage of 0.8 V. The VD, implemented in an experimental chip fabricated in 14-nm CMOS FINFET, exploits set-partitioning principles and embedded per-survivor decision feedback to reduce implementation complexity and power consumption. The active area of the VD with 12 slices, each operating at one-eighth of the modulation rate, is 0.507 x 0.717 mm². Experimental results showing system performance are obtained by using a (2¹⁵ - 1)-bit pseudo-random binary sequence. The impact of the synchronization length and survivor path memory length on the detector design and system performance are shown. A new pipelined reduced-state sequence detector algorithm is presented for high-speed implementations. A novel speculative symbol timing recovery scheme is proposed. New simulation results are obtained to compare the performance of the Reed-Solomon (RS)-encoded 4-PAM scheme with that of the concatenated RS 4-D 5-PAM trellis-coded-modulation (TCM) scheme over an ideal band-limited additive-white-Gaussian-noise channel. Drawing on the results achieved for the VD, novel design techniques for a high-speed low-complexity eight-state 4-D 5-PAM TCM decoder is proposed.
A trade-off between accuracy and fairness is almost taken as a given in the existing literature on fairness in machine learning. Yet, it is not preordained that accuracy should decrease with increased fairness. Novel to this work, we examine fair classification through the lens of mismatched hypothesis testing: trying to find a classifier that distinguishes between two ideal distributions when given two mismatched distributions that are biased. Using Chernoff information, a tool in information theory, we theoretically demonstrate that, contrary to popular belief, there always exist ideal distributions such that optimal fairness and accuracy (with respect to the ideal distributions) are achieved simultaneously: there is no trade-off. Moreover, the same classifier yields the lack of a trade-off with respect to ideal distributions while yielding a trade-off when accuracy is measured with respect to the given (possibly biased) dataset. To complement our main result, we formulate an optimization to find ideal distributions and derive fundamental limits to explain why a trade-off exists on the given biased dataset. We also derive conditions under which active data collection can alleviate the fairness-accuracy trade-off in the real world. Our results lead us to contend that it is problematic to measure accuracy with respect to data that reflects bias, and instead, we should be considering accuracy with respect to ideal, unbiased data.