You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!

  • SRAM Optimization for 14nm and 28nm FDSOI

    I've done SRAM and DRAM design before as a circuit designer from 1978-1986, but in 2016 there are so many more challenges to using 28nm and 14nm on FDSOI technology. One way to keep abreast of SRAM design is to read conference papers, so I just finished a paper from authors at STMicroelectronics and MunEDA presented at the IEEE IRPS 2016 (International Reliability Physics Symposium) held April 17-21 in Pasadena, CA. The paper is titled, BTI Induced Dispersion: Challenges and Opportunities for SRAM Bit Cell Optimization.

    Let's start with a 6 transistor SRAM cell schematic showing a Word Line (WL), Bit Lines (BL), PMOS pull-up transistors (PU), NMOS pull-down transistors (PD) and NMOS transfer gate transistors (PG):

    Article: A Brief History of the Fabless Industry-sram-cell-min.jpg

    BTI
    One reliability challenge is that the Voltage Threshold (Vth) of both PMOS and NMOS transistors will change over time as the devices are turned on and off. For PMOS transistors the effect is called Negative-Bias Temperature Instability (NBTI) and for NMOS transistors its called Positive-Bias Temperature Instability (PBTI). With NBTI and PBTI the value of Vt increases over time and mobility decreases, causing the transistors to become slower.

    Memory Challenges
    Because an SRAM bit cell is very sensitive to device matching we really need to know how the Vth distribution changes over time. The minimum VDD supply level required to ensure proper write and read cycles is called Vddmin, and it's another critical metric for users of an SRAM instance.

    Test Structures
    Engineers at STMicroelectronics designed a test structure in 14nm FDSOI to characterize SRAM bit cells, then measured the difference in MOS values with fresh and aged parameters for both NMOS and PMOS transistors:

    Article: A Brief History of the Fabless Industry-test-structure-measurements-min.jpg
    Correlation plots between fresh and aged parameters

    For PMOS transistors we find that the spread is larger after aging than NMOS transistors, while the parameters are only shifted for NMOS transistors. The next chart shows how the Vth values are shifted for different stress times on the left, and on the right it shows that Vth shift is uncorrelated with the initial Vth.

    Article: A Brief History of the Fabless Industry-vth-shift-min.jpg

    The Vth distribution is still normal as shown below with devices i and j starting out at plus 3 sigma Vth and minus 3 sigma Vth, respectively. The worst Vth after a BTI stress is directly related to its initial Vth.

    Article: A Brief History of the Fabless Industry-vth-normal-distribution-min.jpg

    A BTI model can be developed to show the Vth shift and its final distribution.

    Article: A Brief History of the Fabless Industry-bti-model-min.jpg

    Using this BTI model and then running Monte Carlo analysis on a 6 transistor SRAM bit cell to measure the Static Noise Margin (SNM) during write and read modes produces the next chart, along with a Worst Case Analysis (WCA) using the WiCkeD software from MunEDA:

    Article: A Brief History of the Fabless Industry-static-noise-margin-min.jpg

    Sensitivity analysis of local Vth parameters at a 5 sigma target for SNM reveals which device parameters are the major contributors. The contribution of ageing induced dispersion is moderate compared to the one of fresh Vth location variation, but is significant.

    • Fast pull-down R
    • Slow pull-up R
    • Fast pass-gate R

    Article: A Brief History of the Fabless Industry-sensitivity-analysis-min.jpg

    SRAM Bit Cell Sizing
    For optimal bit cell design we need to optimize SNM, Write Margin (WM), Iread (drive current during read) and Isb (leakage current). Fortunately for us there is an automated Yield Optimizer (YOP) that can attack this problem, instead of using manual, iterative efforts. Here's a screen shot of using the yield optimization from the WiCkeD tool for: Area, Iread, Isb, SNM, WM

    Article: A Brief History of the Fabless Industry-yo-results-min.jpg
    Note how both WM and SNM grow in robustness during optimization, whereas their nominal values grow only little or even shrink. Such mixed effects between spec robustness, nominal value, and device geometries are common and are the reason why a yield optimizer has to run high sigma analysis repeatedly for all specs. Short runtime and high accuracy of the worst-case analysis are key for SRAM yield optimization.

    You can even ask the optimizer to tune transistor sizes for low leakage, low power or high performance.

    Article: A Brief History of the Fabless Industry-different-performance-min.jpg

    Summary
    SRAM design is a tricky thing to use manual efforts and then get optimal results because of physical effects like BTI. STMicroelectronics has worked with MunEDA to create a methodology that automates transistor sizing of their SRAM bit cell while taking into account BTI effects on Vth values. This optimization helps create higher-yielding SRAM IP for use in SoCs built with FDSOI.

    Related Blogs