Transcript Document
An Algorithm for Optimal Decoupling Capacitor Sizing and Placement for Standard Cell Layouts Haihua Su, Sani R. Nassif IBM ARL Sachin S. Sapatnekar ECE Department University of Minnesota 7/17/2015 ISPD'02, San Diego, CA 1 Outline On-chip decap overview Modeling and noise analysis Problem formulation and Adjoint sensitivity analysis Decap sizing and placement scheme Experimental results Conclusion 7/17/2015 ISPD'02, San Diego, CA 2 On-chip Decoupling Capacitors Non-switching gate capacitance Thin oxide capacitance C w: h: tox: ox: 7/17/2015 ox t ox w h width of decap height of decap thickness of thin oxide permittivity of SiO2 ISPD'02, San Diego, CA 3 Decoupling Capacitor Models 1st order model 2nd order model (non-idealities) 7/17/2015 ISPD'02, San Diego, CA 4 Power Network Modeling Power Grid: resistive mesh Cells: time-varying current sources Decaps: 1st order or 2nd order decap model Package: inductance + ideal constant voltage source + 7/17/2015 ISPD'02, San Diego, CA 5 Power Grid Noise Analysis Noise metric: shaded area Vj + z(j) Z = S z(j) Waveform of node j on VDD grid Reference: A. R. Conn, R. A. Haring and C. Visweswariah, Noise Considerations in Circuit Optimization, ICCAD’98 7/17/2015 ISPD'02, San Diego, CA 6 Formulation - Constrained Nonlinear Programming Problem Minimize Z(wj), j = 1..Ndecap Subject to Swk (1-ri)Wchip, i = 1..Nrow And 0 wj wmax , j = 1..Ndecap – ri is the occupancy ratio of row i Cell Decap wj 7/17/2015 ISPD'02, San Diego, CA 7 Solver – Sequential Quadratic Programming (SQP) QPSOL - Quasi-Newton method to solve the problem of multidimensional minimization of functions with derivatives Requirements – – evaluation of the objective function and constraint functions calculation of first-order derivatives 7/17/2015 ISPD'02, San Diego, CA 8 Adjoint Sensitivity Analysis Original circuit Gx(t ) Cx (t ) u(t ) Vj(t) + Adjoint circuit Gxˆ( ) Cxˆ( ) i( ) x(t) and xˆ ( ) i j( ) – node voltages, source currents, inductor currents u(t) – time-dependent sources i() – current sources applied to all bad nodes 7/17/2015 ISPD'02, San Diego, CA 9 Adjoint Sensitivity Analysis (cont’d) Convolve to get sensitivities T Z C (T t )vC (t )dt C 0 T Z R (T t )iR (t )dt R 0 Z is the noise metric for all the grid = S z(j) 7/17/2015 ISPD'02, San Diego, CA 10 Adjoint Sensitivity Analysis (cont’d) Fast convolution for piecewise linear waveforms ~O(N+M) N linear segments q p (a bt)'[ g k (T t )]dt pq b[ g k (T t )]dt b(q p)[ g kT k (q p) / 2] p 7/17/2015 q M linear segments ISPD'02, San Diego, CA 11 Sensitivity w.r.t. Decaps Adjoint sensitivity w.r.t. Cnear, R and Cfar Applying chain rule to find the sensitivity w.r.t. decap width w: Z Z Cnear Z R Z C far w Cnear w R w C far w 7/17/2015 ISPD'02, San Diego, CA 12 Scheme Analyze circuit and store waveforms Compute Z Setup current sources for adjoint circuit Analyze adjoint circuit & store waveforms Compute Z/Ci and Z/wi Evaluate constraint function & gradients Feed to QP solver to get the updated wi According to the new wi , replace cells and decaps one by one 7/17/2015 ISPD'02, San Diego, CA 13 Decap Optimization Process (one row for illustration) Start from equal distribution of decaps: Iteration 1: Iteration 7/17/2015 2: ISPD'02, San Diego, CA 14 Optimization Results Vdd =1.8V, vdrop limit =10%Vdd, ri = 80% Chip Opt Num bad nodes 1 Before After 105 2 974 0.193 0.176 0.121 0.000 53 1964 0.9 2 Before After 80 63 861 0.230 0.196 0.366 0.063 85 3288 15.2 3 Before After 100 70 828 0.222 0.201 0.649 0.200 132 3664 12.5 7/17/2015 Num Vmax of (V) nodes Z (Vns) Num of rows Num of dcps CPU time (mins) ISPD'02, San Diego, CA 15 VDD and GND Contour (chip2) Vmax=0.190V Vmax=0.191V Z=0.366(V•ns) 7/17/2015 Vmax=0.230V Vmax=0.196V Z=0.063(V•ns) ISPD'02, San Diego, CA 16 Optimal Placement (chip2) 7/17/2015 ISPD'02, San Diego, CA 17 Noise Reduction Trend (chip2) 7/17/2015 ISPD'02, San Diego, CA 18 Conclusion Proposed a scheme of decoupling capacitor sizing and placement for standard-cell layouts Applied after placement and before signal routing Formulated into nonlinear programming problem Reduced transient noise Presented a fast piece-wise linear waveform convolution for adjoint sensitivity analysis 7/17/2015 ISPD'02, San Diego, CA 19 Thank you! 7/17/2015 ISPD'02, San Diego, CA 20