Abstract
As CMOS technology continues to scale down, the effect of process variations on yield and performance of analog/RF ICs is becoming more prominent. To counteract this effect, learning-based post-production tuning has been proposed, wherein regression functions are trained and used to adjust tunable knobs based on low-cost alternate tests, thereby improving the performances of a circuit and, by extension, increasing yield. Of course, tunable knobs are also subject to process variations; yet this is not an issue when the knobs are part of the procedure that generates the data with which the regression models are trained, as this data reflects the impact of process variations on both the tunable circuit and the knobs. In various cases, however, such as in heterogeneous integrated systems, 3D ICs, or multi-chip modules, the knob circuitry may not be integrated on the same die, thereby limiting our ability to obtain a comprehensive set of training data. Accordingly, in this work we investigate the impact of knob non-idealities which are not captured in the training data, on the ability of the learned regression functions to accurately predict the optimum knob position that maximizes the performance of a circuit. Using a tunable cascode low-noise amplifier (LNA) fabricated in 130nm CMOS process, alongside external knobs designed as linear low drop out regulators (LDOs) and voltage dividers operating on the bias voltages of the LNA, we first quantify this impact. Then, we demonstrate that by explicitly introducing “noise” in the knob output values used during training set generation, we can effectively alleviate external knob non-idealities and improve quality of tuning.