1 / 19

Design Sensitivities to Variability: Extrapolations and Assessments in Nanometer VLSI

Design Sensitivities to Variability: Extrapolations and Assessments in Nanometer VLSI. Y. Kevin Cao * , Puneet Gupta + , Andrew Kahng + , Dennis Sylvester # , Jie Yang # * UC Berkeley EECS Dept. + UC San Diego ECE Dept. # U. Michigan EECS Dept. Introduction.

Download Presentation

Design Sensitivities to Variability: Extrapolations and Assessments in Nanometer VLSI

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Design Sensitivities to Variability: Extrapolations and Assessments in Nanometer VLSI Y. Kevin Cao*, Puneet Gupta+, Andrew Kahng+, Dennis Sylvester#, Jie Yang# *UC Berkeley EECS Dept. +UC San Diego ECE Dept. #U. Michigan EECS Dept.

  2. Introduction • Why process variability is important now: • Optical lithography: feature size limited by diffraction, starting with 0.35 μm generation • Same patterns look the same? • ITRS: one of the biggest challenges is Lgate control • Affects both circuit’s performance and yield 0.35 μm global line width

  3. Motivation • Necessity of Design Sensitivity Study • Cost of corrections • May be proportional to complexity of masks • Not all corrections are implementable • Previous studies • Lacking quantified projections of variation in scaled technology • Overly simplified circuit topologies and performance models • Ignoring physical correlations

  4. Interpretation of Results • Returns for alternative process improvements measured as • Selling point yield • Selling point: delay point which gives 99.7% parametric yield with all parameters varying nominally • Required guardbanding • 3/mean is taken as the required guardbanding, assuming 99.7% to be the target parametric yield and the delay distribution to be Gaussian • Contributes insights to the impact of process variation through the next two ITRS technology nodes

  5. Correlation Specification • Vth0 = (Tox, Nch, Leff, and Xt) • Perfect correlation between NMOS and PMOS within a gate • Negative correlation between metal width and spacing variation as well as ILD and H variation • Spatial correlation among devices and interconnect • Correlation coefficient linear decays with distance

  6. Comparison with plots • “No correlation coefficient model” is about 6% optimistic • We could lose 14% more from the “perfect correlation model”

  7. Required Guardbanding • The guardbanding value drops from 18.5% for 180nm to 13% for 70nm • Effect of process control • Leff control has the most impact • Intense sensitivity to Vdd Effect of process control on required guardbanding to achieve 99.7% parametric yield

  8. Selling Point Parametric Yield • Individual analysis for the effect of Leff and Vdd control on selling point parametric yield • Loose control of Leff (Vdd) can cause loss up to 5% (2%) in yield

  9. No Further Control • Effects of no further investments for control of Leff and Vdd compared with 180nm technology node • Results confirm that performance sensitivity to Leffand Vddvariation is high

  10. Conclusions • More accurate results with realistic correlation among variability included • With qualified projection of variation in scaled technology, the impact of variability is actually decreasing • Performance very sensitive to Leff variation • Considering the huge cost of Leff control, tackling Leff variation from design perspective may be more cost-effective • Vdd is an important source of variation • The control of Vdd variation may give a better ROI than Leff control for achieving the same amount of variability–tolerance as technology scales

  11. Ongoing Research • Incorporate multiple correlated critical paths • More trials in MC simulation for more accurate results • Set up an analytical flow to parallel testbed of MC simulation using HSPICE • Set up cost-driven correction flow based on the previous results • Relax specs on some (non-critical) gates • Goal: reduce costs with zero parametric yield penalty

  12. Toward Performance-Driven Reduction of the Cost of RET-Based Lithography Control Dennis Sylvester (dennis@eecs.umich.edu), Jie Yang (Univ. of Michigan, Ann Arbor) Puneet Gupta, Andrew B. Kahng (UC San Diego)

  13. Trends in Mask Cost.. • RETs increase mask feature complexities and hence mask costs • The average mask set produces only 570 wafers  amortization of mask cost is difficult • Mask writers work equally hard to perfect critical and non-critical shapes; errors found in either during mask inspection will cause the mask to be discarded

  14. MinCorr: The Cost of Correction Problem.. • Define the selling point as the circuit delay which achieves 99% parametric yield The MinCorr problem seeks a level of correction for each layout feature such that a prescribed selling point delay is attained with minimum total cost of corrections.

  15. MinCorr: Parallels to Gate Sizing.. • Use off-the-shelf synthesis tool, along with yield library similar to timing libraries (e.g., .lib) to perform OPC “sizing” operation

  16. MinCorr: Yield Aware Library Characterization • Mask cost is assumed proportional to number of layout features • Monte-Carlo simulations, coupled with linear interpolation, are used to estimate delay variance given the CD variation • We generate a library similar to Synopsys .lib with (+3) delay values for various output loads

  17. Experiments and Results • Three levels of OPC considered • Input slew dependence ignored • Interconnect variation ignored Sample Result of Library Generation

  18. Experiments and Results • 4 combinational testcases ranging from 1600 to 9400 gates Sample results on 9410 gate testcase alu128

  19. Experiments and Results • Small (4%) selling point delay variation between max- and min-corrected versions of design • Sizing-based optimization achieves 60-79% reduction in OPC cost without sacrificing parametric yield

More Related