Fang Gong

The reliability issue of highly replicated circuits has become increasingly challenging under large-scale process variations, because the circuit failure is a "rare event" with an extremely small probability (e.g., 1e-5~1e-11). Among existing techniques, importance sampling is a very widespread solution providing high efficiency. However, the "curse of high dimensionality" prevents it to be used in high dimensions, therefore, making it unsuitable for real problems in practice.

In this work, we proposed a novel and efficient statistical analysis of rare events in high dimension, which has been successfully applied to predict the extremely small failure probability of memory circuits. To the best of our knowledge, this is the first work that applies the importance sampling paradigm to high dimensional problems (e.g., tens or hundreds random variables) . The experiments on an memory circuits (e.g., SRAM bit-cell, delay chain) show the proposed algorithm can provide several orders of magnitude speedup over Monte Carlo method without compromising any accuracy. In addition, the proposed approach is tens times faster than classification based method (e.g., Statistical Blockade) whereas existing importance sampling based methods completely fail to provide reasonable accuracy.

This work is presented in DAC 2012 WIP [O3] and is under review.

This work tries to estimate the failure probability of rare-failure event in SRAM cell, which is extremely small. We propose to improve the conventional importance sampling method by probability collectives.

First, we use the "Kullback-Leibler distance" (KL) to measure the distance between the optimal sampling distributions and the given sampling distributions of variable process parameters. Then, the probability collectives technique is adapted to analytically minimize this KL distance so as to find a sampling distribution as close to the optimal as possible.

The proposed algorithm significantly expedite the convergence of importance sampling method. Experiments demonstrate proposed algorithm is 5200X faster than Monte Carlo method and achieves more than 40X speedup over other existing state-of-the-art techniques without compromising any accuracy.

This work has been published by ISPD 2012 [C7]. The source code can be download from here.

The random device noise has become one of fundamental limits for analog/RF circuits. To describe the effects of device noise in time domain efficiently, this work apply the recent-established polynomial chaos to the stochastic differential algebraic equation (SDAE). As such, the noise variance along the time can be extracted with only one-time calculation. It considers both thermal noise and flicker noise in the transient noise analysis, which can be 488X faster than Monte Carlo method with similar accuracy and achieves on average 6.8X speedup over existing Non-Monte-Carlo approach.

This work has been published by the 48th ACM/IEEE DAC 2011 for publication C[6].

This work builds a low-cost ($19) and wearable (tens grams) BCI system to acquire Electroencephalography (EEG) signals. The system can detect user's concentration based on EEG signal analysis in real-time manner. It can further control the surrounding appliances such as reading lamps, fans and etc. Potential applications include gaming, wheelchair motion control for disabled, and brain health monitoring.

       
(a) prototype circuit of portable BCI system                         (b) packaged smart headset of BCI system

This work has been demoed in the third UCLA Technology and Aging Conference in Oct 2010. Visitors to the conferences, in large numbers, lined-up to try this system, amazed at the BCI's functionality. Also, it has been published in [J6] and reported in BIT and New Media Journal.

Collaborators: Wenyao Xu, Ju-Yueh Lee, Prof. Majid Sarrafzadeh

This work explores the efficient extraction of the variational capacitance in presence of process variations. It first defines geometric moments, which relates the potential coefficients with geometry information, and then develops one parallel fast multi-pole algorithm based on those moments. In addition, it presents spectrum pre-conditioner to consider different variation sources incrementally. The entire extraction flow is called "PiCAP", which can achieve 15X speedup over serial counterpart.

This work was published in the 46th ACM/IEEE DAC 2009 [C2] and TVLSI [J3].

This work develops an efficient field solver (named "Mixed Boundary Element Method" or "MBEM" ) to extract the wideband impedance from zero frequency to at least tens of Giga Hertz, which can be applied to complicated 3D structures and obtains great advantage over FastImp (M.I.T developed in 2005), on both accuracy and efficiency. It utilizes the efficient pFFT algorithm (pre-corrected Fast Fourier Transform) to speed up the evaluation of matrix-vector-product (MVP) in iterative solvers such as Generalized Minimal Residual Method (GMRES). Specifically, the complexity of MVP can be reduced from O(N²) to nearly O(N), where N is the number of variables. Moreover, this work presents many numerical optimization skills to avoid redundant operations in order to further improve the efficiency.

This work has been published with [J1], and on the 13th ASPDAC’08 [C1] .