Intel (Israel)

The celebrated Algorand protocol solves validated byzantine agreement in a scalable manner in the synchronous setting. In this paper, we study the feasibility of similar solutions in the asynchronous setting. Our main result is an asynchronous validated byzantine agreement protocol that we call Asynchronous Algorand. As with Algorand, it terminates in an expected constant number of rounds, and honest parties send an expected O(n polylog n) bits, where n is the number of parties. The protocol is resilient to a fully-asynchronous weak-adaptive adversary that can corrupt a near-optimal number of parties (< (1/3 - ϵ)n) and requires just a verifiable random function (VRF) setup and secure erasures.

This paper presents a power amplifier (PA) in the 76-81 GHz automotive radar band, implemented in an advanced 16 nm FinFET CMOS node. The PA consists of 3 common-source (CS) neutralized differential pairs (NDP) and transformer-based matching networks. The first two stages allow digital gain and current consumption control. The last stage utilizes power combining to boost the output power while adhering to the stringent reliability constraints. The PA delivers a measured saturated output power of 13.2-14 dBm across the 76-81 GHz band. It achieves measured small signal gain in full gain mode of 30.7-32.3 dB in that band and gain tuning range of 33 dB. The PA occupies core area of 460 µm x 150 µm and consumes 130 mA from 0.95 V supply in saturation.

For the most advanced nodes, edge placement errors are typically dominated by stochastics, necessitating a rigorous stochastics approach to modeling and measuring edge placement errors and their contributors. Recently, a new approach to developing an edge placement error (EPE) model useful for lot dispositioning or EPE budgeting was presented. [1] This model gave rise to the generation of an Overlay Process Window (Figure 1), the range of overlay errors that can be tolerated in the presence of stochastics critical dimension and placement errors of the individual layers. In this work, the new EPE modeling approach is applied to other uses cases, including a contact hole landing on a line. Further, a propagation of uncertainty analysis is applied to all input measurements in order to determine the metrology requirements for measurements of the stochastics variables such as local CD uniformity and local pattern placement error. The result will be a wider ranging and more useful mode

A 224-Gb/s pulse amplitude modulation 4-level (PAM4) ADC-based SerDes receiver (RX) is implemented in a 5-nm FinFET process. The RX consists of a low-noise hybrid analog front-end (AFE) that incorporates both inductive peaking and source degeneration, a 64-way time-interleaved ADC, digital equalization consisting of an up to 30 tap feed-forward equalizer (FFE), optional decision-feedback equalizer (DFE), and a clock-data-recovery (CDR) loop utilizing a 14-GHz digitally controlled oscillator (DCO). The RX was characterized for medium-to-long reach channels (20.6-, 27-, 31.4-, and 38-dB insertion loss at Nyquist) with a corresponding pre-forward error correction (FEC) bit error rates of 1.9E-11, 5.2E-10, 1.2E-8, and 6E-7, respectively. The analog power consumption of the RX is 1.41 pJ/b.

Abstract As a subfield of artificial intelligence, machine learning designed to learn the structure of the data. Machine learning has been widely used in many scientific problems. In this study, we used machine learning techniques to figure out the most important physicochemical properties for type classification of red wines. We used a wines' dataset with 13 physicochemical properties. We used a Random Forest classifier to predict wine's type from its features, and permutation feature importance, in order to detect the most important properties of the wine for type classification. The properties: flavanoids, proline, and color intensity were found to be most important for type classification. Additional 4 classifiers: Laso classifier, Ridge classifier, Decision Tree classifier, and Support Vector classifier were used and examined for classification and feature importance. Flavanoids and proline were very important across all classifiers.

A key element in solving real-life data science problems is selecting the types of models to use. Tree ensemble models (such as XGBoost) are usually recommended for classification and regression problems with tabular data. However, several deep learning models for tabular data have recently been proposed, claiming to outperform XGBoost for some use cases. This paper explores whether these deep models should be a recommended option for tabular data by rigorously comparing the new deep models to XGBoost on various datasets. In addition to systematically comparing their performance, we consider the tuning and computation they require. Our study shows that XGBoost outperforms these deep models across the datasets, including the datasets used in the papers that proposed the deep models. We also demonstrate that XGBoost requires much less tuning. On the positive side, we show that an ensemble of deep models and XGBoost performs better on these datasets than XGBoost alone.

This work presents a divide-by-2 frequency divider designed using 22-nm CMOS Fully-Depleted Silicon on Insulator (FDSOI) technology. The circuit combines a dynamic load modulation, current reuse technique and back gate tuning to obtain a division range of 10 to 92 GHz with 0dBm input signal to the divider core (due to measurement limitations measured range is 20 to 60 GHz). The divider is suitable for all 5G bands, WiGiG, E-band radar and others and operates from a CMOS rail with no AC coupling required. The divider consumes less than 2 mW from 1V supply to obtain a figure-of-merit of more than 39 GHz/mW and occupies less than 65 um <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of silicon area.

We point out that common evaluation practices for cross-document coreference resolution have been unrealistically permissive in their assumed settings, yielding inflated results. We propose addressing this issue via two evaluation methodology principles. First, as in other tasks, models should be evaluated on predicted mentions rather than on gold mentions. Doing this raises a subtle issue regarding singleton coreference clusters, which we address by decoupling the evaluation of mention detection from that of coreference linking. Second, we argue that models should not exploit the synthetic topic structure of the standard ECB+ dataset, forcing models to confront the lexical ambiguity challenge, as intended by the dataset creators. We demonstrate empirically the drastic impact of our more realistic evaluation principles on a competitive model, yielding a score which is 33 F1 lower compared to evaluating by prior lenient practices.

The goal of this study is to improve the scalability of today's SAT-based solutions for optimization problems and to pave the way towards extending the range of optimization problems solvable with SAT in practice. Let OptSAT be the problem of optimizing a generic Pseudo-Boolean function, given a satisfiable propositional formula F. We introduce an incremental and anytime incomplete algorithm for solving OptSAT, called Polosat. We show that integrating Polosat into a state-of-the-art open-source anytime MaxSAT solver significantly improves the solver's performance. Furthermore, we demonstrate that Polosat substantially improves the solution quality of an industrial placement tool, where placement is a sub-stage of the physical design stage of chip design.