Post-Silicon Receiver Equalization Metamodeling by Artificial Neural Networks

Abstract

As microprocessor design scales to the 10 nm technology and beyond, traditional pre-and post-silicon validation techniques are unsuitable to get a full system functional coverage. Physical complexity and extreme technology process variations severely limits the effectiveness and reliability of presilicon validation techniques. This scenario imposes the need of sophisticated post-silicon validation approaches to consider complex electromagnetic phenomena and large manufacturing fluctuations observed in actual physical platforms. One of the major challenges in electrical validation of high-speed input/output (HSIO) links in modern computer platforms lies in the physical layer (PHY) tuning process, where equalization techniques are used to cancel undesired effects induced by the channels. Current industrial practices for PHY tuning in HSIO links are very time consuming since they require massive lab measurements. An alternative is to use machine learning techniques to model the PHY, and then perform equalization using the resultant surrogate model. In this paper, a metamodeling approach based on neural networks is proposed to efficiently simulate the effects of a receiver equalizer PHY tuning settings. We use several design of experiments techniques to find a neural model capable of approximating the real system behavior without requiring a large amount of actual measurements. We evaluate the models performance by comparing with measured responses on a real server HSIO link.

Transmitter and Receiver Equalizers Optimization Methodologies for High-Speed Links in Industrial Computer Platforms Post-Silicon Validation

“…In [25], we propose a metamodeling approach, based on artificial neural networks (ANN), to efficiently simulate the silicon Rx equalizer. The model is generated using a frugal set of training data exploiting several DoE approaches to reduce the number of test cases.…”

“…The 3LP is trained by using the Bayesian regularization [29] method available in MATLAB Neural Network Toolbox. The algorithm for training the ANN is shown in [25]. We first define the learning ratio to split the pairs of inputs and targets into the learning and testing datasets.…”

“…We start training the 3LP with just one hidden neuron (h = 1), and calculate the corresponding learning and testing errors. We keep increasing the complexity of the ANN (h) until the generalization performance starts to deteriorate [25].…”

“…This paper presents several optimization techniques based on novel objective functions to optimize the transmitter (Tx) and receiver (Rx) equalizers in a server post-silicon validation platform. This paper essentially summarizes [2], [11] and [25].…”

Advanced Neural-Based Model Generation and Extrapolation Techniques for Microwave Applications

“…The sparsity of this matrix D 11 is 97.02%, confirming that D 11 is a large sparse matrix. design specifications of the bandstop filter defined as follows[89] |S21 | 0.9, for 5 GHz ω 8 GHz |S 21 | 0.05, for 9.35 GHz ω 10.75 GHz |S 21 | 0.9, for 12 GHz ω 15 GHz…”