Somha, Worawit

Loading...
Thumbnail Image
Preferred name
Somha, Worawit
Main Affiliation
Email
worawit.so@kmitl.ac.th

Settings

Now showing 1 - 2 of 2
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    A segmentation kernel fitting technique to circumvent extreme deviation from exponentially descent tail distribution
    (2018-07-01)
    Somha, Worawit
    ;
    Yamauchi, Hiroyuki
    A segmentation kernel fitting technique has been proposed to circumvent an extreme deviation from the exponentially steeping descent tail distribution in the deconvolution. The proposed technique regenerates each segmented distribution line by finding the minimum of unconstrained multivariable function using derivative-free method. We decomposed the convolution effects of the two types of the minimum operating voltage variations caused by the spatially random threshold variation (VDD<inf>SPAT</inf>) and the temporally random threshold variation (VDD<inf>TIME</inf>), respectively. We discussed the VDD<inf>SPAT</inf> and VDD<inf>TIME</inf> effects on the SRAM fail-bit count (FBC) based on the decomposing results. It is found that the FBC estimation error for the proposed one can be reduced to almost 14-orders of magnitude smaller than that for the off-the-shell functions.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    A dual-band filter designed for retrieving both left- and right-tailed RTN distributions in an iterative deconvolution procedure
    (2025-05-01)
    Yamauchi, Hiroyuki
    ;
    Somha, Worawit
    This paper proposes a dual-band filter design to alleviate a crucial ringing issue of the Richardson–Lucy deconvolution algorithm (RL-deconv). We found that the RL-deconv doesn’t work for an exponentially decaying tail due to the ringing. The reasons why we must handle this issue in the VLSI chip reliability design are: (1) the VLSI chip bit density has increased up to a 10<sup>12</sup>-bit scale, making the fail probability obey the long tail down to 10<sup>–12</sup>, and (2) the VLSI chip margin variations have become prominent, caused by atomic-level random behaviors. The tail for the variations caused margin variations to obey the Gamma distributions. Consequently, the tail of the VLSI chip margin distribution doesn’t follow the Gaussian distribution anymore. These backgrounds compel us to newly adopt the inverse problem methods to predict the tail distribution based on the deconvolution. As the tail gets longer, the element-wise misalignment becomes larger between the corresponding elements of the feedback gain and the objective of deconvolution, and this causes to more critical wrong element-wise amplification, leading to a ringing. We found that it is not sufficient to retrieve only the right tail because the left tail can no longer be ignored. The length of the left tail becomes long enough to influence the distribution after aging. To address this issue, this paper proposes a dual-band filter design for retrieving both left and right tails, which contributes to widening the alignment range of the feedback gain with the retrieving target in the RL-deconv iterative processes. It is found that the proposed technique reduces the RTN deconvolution error by 12-fold compared with the conventional one.