Abstract—This paper presents a comparative analysis of the Random Telegraph Noise (RTN) deconvolution accuracy between the Richardson-Lucy (R-L) algorithm and the proposed partitioned forward problem based deconvolution means (PFDCV). Unlike the R-L based deconvolution, the proposed technique successfully solves the issue of noise amplification thanks to eliminating any operations of differential and division. This effectiveness has been demonstrated for the first time with applying it to a real analysis for the effects of the RTN on the overall SRAM margin variations. It has been shown that the proposed PFDCV technique can reduce its relative errors of the RTN deconvolution by 10 14 –fold compared with the cases of the R-L. Index Terms—Random telegraph noise, fail-bit analysis, static random access memory, deconvolution, richardson-lucy deconvolution. I. INTRODUCTION The approximation-error of the tails of random telegraph noise (RTN) distribution will become a crucial challenge. This stems from the facts that: (1) tails of the RTN distribution (g) will become longer than that of random-dopant-fluctuation (RDF) (f) that is previously dominant factor of overall margin-variations, as shown in Fig. 1 and (2) the convolution result (h=fg) of the RDF(f) with the RTN(g) will be more governed by the RTN than the RDF, as can be seen in the comparison of (h=fg) between Fig. 2(a) and Fig. 2(b) for short and long RTN, respectively. Because the increasing paces of variation-amplitude Vth of the threshold voltage (Vth) are differently dependent on the MOSFET channel-size (LW) like the below expressions of (1) and (2), the Vth increasing paces of the RTN is a 1.4x faster than that of the RDF if assuming the LW is scaled down by 0.5-fold every process generation, as shown in Fig. 1. ( ) ( ) Vth RDF AVt RDF LW   (1) ( ) ( ) Vth RTN AVt RTN LW   (2) where AVt (RDF) and AVt (RTN) are Pelgrom coefficients for the RDF and the RTN, respectively. According to the references [1]-[4], there will come the time soon around a 15nm-scaled CMOS era. Manuscript received December 5, 2013; revised February 26, 2014. This work was supported in part by MEXT/JSPS KAKENHI Grant Number of 23560424 and grant from Information Sceience Laboratory of Fukuoka Institute of Technology. The authors are with the Information Intelligent System Fukuoka Institute of Technology, 3-30-1, Wajiro-Higashi, Higashi-ku, Fukuoka, Japan (e-mail: bd12002@ bene.fit.ac.jp, yamauchi@fit.ac.jp). Fig. 1. Trend of variation amplitude of RTN and RDF. Variation amplitude of RTN becomes larger than that for RDF in 10nm era. Fig. 2. Comparison of the convolution result h=fg of the RDF(f) with RTN(g) between (a) short RTN and (b) long RTN. Convolution result h becomes governed by the RTN(g) when the tail length of g is larger than that for f. The reliability design for the static random access memory (SRAM) will become an unprecedentedly crucial challenge because the increased time-dependent (TD) margin variations (MV)-caused failures cannot be predicated any more by only ordinary convolution analyses [1]-[4]. This stems from the facts that latent TD-MV, (i.e., unknown MV after shipped to the market), will become much larger than already-known MV based on the measurements in advance. This leads to an increased pressure to figure out the unknown factors by solving the inverse problem [5]-[9], although the SRAM designers are unfamiliar with such kind of methodology until now. Fig. 3(a) and Fig. 3(b) show an example for the deconvolution ( -1 ) and the convolution (), respectively. Where  -1 and  are arithmetic symbol for deconvolution and convolution, respectively. Fig. 3(a) recounts the following scenarios: a certain distribution (h) within the product target spec (SP prod ) is predefined and the RDF distribution (f) is already-known based on the measured data. The f is truncated at a certain point (TP) based on the screening spec and converted to f TP . However, the TP of the f TP and the random telegraph noise A Comparative Study on RTN Deconvolution of Richardson-Lucy and Proposed Partitioned Means for Analyzing SRAM Fail-Bit Prediction Accuracy Worawit Somha and Hiroyuki Yamauchi International Journal of Computer and Communication Engineering, Vol. 3, No. 3, May 2014 178 DOI: 10.7763/IJCCE.2014.V3.315