Next-Generation Compact Modeling* Robert W. Dutton and Chang-Hoon Choi Stanford University Motivation--Moore's Law scaling has led to ultra-short channel length devices that, while giving multi-GHz performance, present a host of new modeling challenges, especially for analog devices in SoC integration. There are a range of "other" issues that face compact modeling of nanometer scale technology, including parasitic effects related to: gate leakage, substrate coupling thermal limitations There are also issues of intrinsic device scaling; high-level language (HLL) specifications for models that facilitate model portability is one example. *Support from SRC(TI); CIS(Infineon, Philips); MARCO(MSD) Outline* • Intrinsic Devices: – Challenges of MOS Scaling • Gate Current, QM Effects… (C.-H. Choi, PhD 2002) – RF Modeling • Non-Quasi-Static & Substrate Effects (J. Jang PhD 2004) • Thermal Noise (T. Oh, PhD 2004) – Thermal Modeling • Self Heating in Nano-Devices (E. Pop) • RF Power MOS (C. Ito) • Model Portability • Summary and Discussion *PhD Theses and related publications: www-tcad.stanford.edu Scaling Challenges of CMOS silicide side- Poly Gate sidewall wall extrinsic SDE o channel h SDE extrinsic alo l implant ha drain source p-well implant substrate Cross-section of a 90nm N-channel MOS transistor, including details of gate, sidewalls and substrate doping profiles (SDE, channel/well implants, halo doping) Threshold, Sub-Threshold & Ion/Ioff Drain Current (A/um) 10-3 10-4 10-5 10-6 10-7 10-8 10-9 10-10 10-11 10-12 10-13 10-14 Vds=3.0V L=0.1µm =50mV Ion Ion/Ioff “ratio” (performance parameter) L=1.2µm L=0.25µm Ioff sub-threshold slope (m) -0.5 0.0 0.5 1.0 tox=4nm 1.5 2.0 Gate Voltage (V) Semi-Log plot of Drain Current (Ids) versus Gate Voltage (Vgs) with Drain Voltage (Vds) as a parameter for three channel length NMOS devices. Mapping Device Physics to Compact Models Gate Effects: CG, IG, QM Source Substrate Effects: CA, CSW RS Gate intrinsic channel Intrinsic MOS: Cox, Vth, µ, νsat Drain Bulk Body Effects: Vth(Vbs), Sub-Vth Bulk Currents: DIBL, BTBT Intrinsic MOS, Gate, Body/Bulk and Parasitic Substrate Capacitance/Resistance Effects Channel Charge and Carrier Mobility! Ichannel ∝ charge ∝ Cox (Vgs - Vth) * carrier velocity * Elateral µchannel εGate Cox = xGate x Depletion (Vgs ) 1 1 1 1 = + = + CG (Vgs ) Cox εSilicon Cox CDepletion (Vgs ) 1 µchannel 1 = ∑ ith scattering µi mechanisms First-order model of how MOS inversion layer channel current (Ichannel) depends of gate-induced charge and carrier velocity. The physical parameters are CG and µchannel. € Scaling of MOS Performance Relative Device Performance 100 Conventional Bulk CMOS SOI <1.2X Bulk Strained Si/SiGe <1.6X Bulk Double-Gate <1.8X Bulk 10 Lithographyenabled progress, exponential Device/materialenabled progress, incremental 1 | | | | | | | | | | | | | 1984 1988 1992 1996 2000 2004 2008 2012 (Year) Mobility Degradation for High-k Gate Stacks • Various mechanisms responsible for degradation: – Remote polar optical phonon scattering – Remote charge scattering – Remote surface roughness – Phase separation • Universal mobility curve – Effective mob. dep. on effective field; good for acoustic phonon scatt. and surface roughness scatt. – Deviation at low inversion due to Coulomb scatt. for SiO2 – More severe deviation for high-k • Aggravated surface roughness • Stronger Coulombic scatt. at low and medium inversion + Gate Tunneling… Saito et. al IEDM 2003 Poly-Gate and QM Effects Ideal C-V (see equation in previous slide) measured CV classical model QM correction model 180 160 poly depletion Cg (pF) 140 120 QM effects 100 80 60 40 20 0 -4 -3 -2 -1 0 1 2 3 Vg (V) Measured and simulated curves of MOS gate capacitance vs. gate voltage for NMOS, poly-silicon gate, 2nm oxide thickness. Curves compare impact of QM effects QM Effects on Channel Charge classical electron distribution E le c tr o n en e rgy (m e V) Si surface electron distribution for QM model in E0 state 20 E’0 conduction band E1 0 E0 40 80 120 Distance from surface (Å) -20 Surface region of bulk MOS device that shows: well potential created by the conduction band energy (- * -); discrete energy levels imposed by QM (E0, E1…); electron distribution associated with E0 level; electron distribution for classical theory (- - -) QM “Poly Depletion” e pile-up QM model Classic electrons n=2x10 20 /cm 3 quantum-well n~0 n-poly ox. P-Si Electron distributions in the n-type poly-silicon gate of an NMOS transistor, based on classical theory versus that for a quantum-based (QM) solution for charge. The peak region shows a “QM depletion” resulting in Vth and C-V shifts Implications of QM Poly Depletion (Fully Depleted Double-Gate-SOI) Poly-gate Poly-gate 2.8x1020/cm3 2.2x1020/cm3 2.5x1020 1.9x1020 2.2x1020 1.0x1020 1.0x1020 <1.0x1019 <1.0x1019 S Electron contour for Vd=0V D S D Electron contour for Vd=1V 2D simulations of QM poly depletion effects for two drain bias conditions. Results show significant gate depletion (left-side) and lateral effects that influence drain-induced barrier lowering (right-side) Threshold and Sub-Threshold Effects 0.01 L=20nm, tox=1nm DG SOI 1E-3 0.45 1E-4 Cg (fF/um) Ids (A/um) 0.50 0.40 0.35 0.30 0.25 with poly QM effects without poly QM 0.20 0.15 -1.0 -0.5 0.0 0.5 Vg (V) 1.0 1.5 with poly QM Vds=1V 0.1V 1E-5 1E-6 1E-7 1E-8 w/o poly QM effects 1E-9 1E-10 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0 Vgs (V) C-V and I-V characteristics for a 20nm channel length DG SOI transistor, comparing QM effects in the poly gate region with idealized (metal-like) gate. Gate Current (Tunneling) vs. tox 2 ) + 0 Gate Current (A/cm 10 n diode source grounded 12.5Å 15Å 10 -2 10 -4 10 -6 18.2Å Jp 10 -8 -1 -0.5 0 0.5 1 1.5 Gate Bias (V) Simulated and measured data for gate current in an MOS capacitor as a function of gate oxide thickness using NEMO; data provided by Hewlett-Packard. Serious Limits to Scaling tox 160 symbols: measured lines: modeling 140 tox=1.8nm Cg (pF) 120 100 80 tox=1.5nm 60 40 20 0 -3 tox=1.3nm -2 -1 0 1 2 Vg (V) Measured and equivalent circuit simulations of imaginary component of input admittance, large area MOS capacitor structure with oxide thickness as parameter. Drain Current with Isub • Distorted drain current (Id) from the real drain current (Id0) due to gate current (Ig): Id0=Id+0.5Ig (Zeitzoff, EDL’03) • But, note that Id0=Id+0.5(Ig-Isub) in the presence of Isub Correction of Drain Current • Over-estimation of corrected drain current (Id0) for Id0=Id+0.5Ig expression • New expression:Id0=Id+0.5(Ig-Isub) in the presence of Isub Correction of Charge-Pumping Current • Measurement of true inversion charge (Ninv) based on device symmetry (Keber, VLSI tech.’03) • Elimination of Ntun from Ninv for high-K or thin oxide Inversion charge pumping Ninv=2NICP_b-NICP_a Modified Ninv Ninv=2NICP_b-NICP_a-Ntun Where DC values of Ntun are used (Masson et. al 1999) Small-Signal Modeling of RF MOSFET Jaejune Jang PhD 2004 Center for Integrated Systems Stanford University Non-Quasi-Static Effect Quasi-Static Non-Quasi-Static Source: Ronald van Langevelde, Philips Research Laboratories • Quasi-Static is when each terminal responds instantaneously to applied signal • If switching time of vg is compatible with transit time of inversion charge, QS approximation fails NQS Effect on Cggeff and Gggeff VG = 3V, VD = 0 V Cggeff 1 10 Gggeff 0 fNQS 0.9 0.8 ) z H 0 1 @ C (g xg a m / g 0.7 fNQS 10 -5 10 -10 10 -15 10 -20 ~f2 0.6 0.5 | g gg | L=0.3um L=0.6um L=1.2um L=2.5um L=10um L=50um 0.4 g C 0.3 0.2 0.1 0 5 10 10 6 10 7 8 10 Frequency (Hz) 10 9 10 10 L=0.3um L=0.6um L=1.2um L=2.5um L=10um L=50um 10 2 10 4 6 10 Frequency (Hz) 10 8 10 10 • Long channel device falls off at lower frequencies • Gggeff ~ f2: directly proportional to gate induced noise • fNQS is defined as a frequency when NQS start Bias Dependency of fNQS VDS = 0 Leff_gs = Leff_gd fNQS(Cgs) = fNQS(Cgd) Leff _ gs ≈ L ⋅ Leff _ gd ≈ L ⋅ C gs C gs + C gd C gd C gs + C gd VDS > 0 Leff_gs > Leff_gd fNQS(Cgs) << fNQS(Cgd) f NQS µ (VGS − VT ) = n 2 2πLeff f NQS vsat = n 2 π L eff Long Channel Short Channel • Leff is determined by charge sharing b/w source and drain • Bias dependency of fNQS is primarily function of Leff , mobility (µ) and, saturation velocity (vsat) ygs and ygd G Cgs Cgd ymvgs rgd S -ysd ymbvbs D -ybs ymxvgb -ygb rgd rgs • ygs and ygd can be represented as series RC • Cgsrgs/Cgdrgd determines fNQS -ybd VGS B VDS L = 1.2 um (measured) Capacitance vs. Channel Length @ 1MHz L = 10 um 0.45 L = 0.3 um 20 0.4 10 0.35 0 0.3 y x C ) -10 V ( Cg -20 / g C ss C sd C sg C 0.15 C ss C sd C sg C Cxx/Cgg Cxx/Cgg ) V 0.25 ( Cg 0.2 / g sb y x sb C -30 0.1 -40 0.05 -50 0 -0.05 0 0.5 1 1.5 V (V) GS V GS (V) 2 2.5 3 -60 0 0.5 1 V 1.5 (V) 2 2.5 3 GS VGS (V) • All 16 capacitances agree well with charge-based capacitances for long channel device • Short channel capacitances show quite different behavior • This is due to existence of substrate resistance (Rsub) Impact of Rsub on Terminal Admittance G Cgs (gm-jwCm)vgs Cgd Csdeff Rsub g mb = Csd + Cbd 2 2 2 1 + ω RsubCbb g sdeff 2 ω 2 Rsub g mb = g sd + CbbCbd 2 2 2 1 + ω RsubCbb gsd S D Csd gmbvbs Cgb Cbs B Rsub ib B’ Cbd As L decreases, Rsub and gmb increase • Rsub is amplified through gmb • Impacts following y-parameters (ydd, yds, ydb, ydg, yss, ysg, ysd, and ysb ) Bias Dependency of Substrate Network Saturation Region vd vs rs_d Cs_s Cs_d rs_s • It’s been reported that substrate network is bias independent not true • Cs_s > Cs_d & rs_s < rs_d in saturation region - Surface area of effective electrode is bias dependent due to charge sharing ybs & ybd G rgs ymvgs rgd • Csub and Rsub is a strong function of bias • Cbs and Cbd are also strongly bias dependent S -ysd ymbvbs D Cbs Cgs Cgd Cbd Cbs rs_s Cs_s -ygb B rs_d Cbd Cs_d VGS VDS Noise Analysis of deepsubmicron MOSFETs Tae-young Oh PhD 2004 Center for Integrated Systems Stanford University MOS Channel Length and Excess Noise Higher speed devices offer great promise for RF Noise Parameters (no unit) However, the changes in intrinsic noise of scaled MOS devices is not clearly understood 7 Uncertain! • The amount and source for this excess noise were still uncertain 6 5 4 • High electric field in short channel MOS device should have relation with this Drain Noise excess noise - Simulation has to handle this carefully (Drift-Diffusion vs. Hydro-Dynamic models) Gate Noise 3 2 1 0.1 1 Channel Length µm) ( 10 Channel Length vs. Drain Noise Parameters PMOS Transistor Drain Noise Parameter Drain Noise Parameter NMOS Transistor 3.2 3.2 Hydrodynamic 2.5GHz Hydrodynamic 1MHz Drift-Diffusion 2.5GHz Drift-Diffusion 1MHz 2.8 2.4 2.0 2.4 •Noise increases with reduced L 2.0 •Modeling of Carrier Transport 1.6 very1.2important 1.6 1.2 0.8 0.8 Classical Limit 0.4 (Van der Ziel) 0.4 0.0 Hydrodynamic 2.5GHz Hydrodynamic 1MHz Drift-Diffusion 2.5GHz Drift-Diffusion 1MHz 2.8 0.0 0.1 1 10 ChannelLength Lengthµm) ( Channel (µm) Bias Condition : Vg =1.08 V, Vd=1.2 V 0.1 1 10 Channel Lengthµm) ( Bias Condition : Vg =-1.07 V, Vd=-1.2 V Measurement vs. Modeling--0.18µm /Hz) Drain Current Noise (A 0.18 µm Channel Length MOS--Measurement and Simulation at 5 GHz (Data from Philips (Scholten, IEDM 02)) 1.0x10-23 2 2 Gate Current Noise (A /Hz) 10-20 10-21 10-22 10-23 0.2 Measured Hydrodynamic 0.4 0.6 0.8 1.0 1.2 Vgs(V) 1.4 1.6 Measured Hydrodynamic 1.0x10-24 1.8 0.4 0.6 0.8 1.0 1.2 Vgs(V) 1.4 1.6 1.8 Two-Lump Impedance Field Model L’ xL’ Linear Saturation Current Injection iin RS 1/gm 2 iout - Sid = ∇Ak Sin gmVgs RL Vgs + € First-Order View of Impedance Field L xL’ L’(1-x) L’ A(x) = RS RS + € 1 gm = x 1− x x 1− x Space Charge x ≈ ≈x + 1− x x + (1− x) Comparison of A(x)--1µm vs 0.18µm L=1µm A (no unit) 1.0 A (no unit) 1.0 0.8 0.8 0.6 L= 0.18µm 0.6 0.4 0.4 0.2 0.0 -0.6 Hydronynamic Drift-Diffusion -0.4 -0.2 0.0 0.2 Positionµm) ( 0.4 0.6 0.2 Hydronynamic Drift-Diffusion 0.0 -0.15 -0.10 -0.05 0.00 0.05 Positionµm) ( •Effects on noise in short-channel devices more pronounced •Field-effects impact source-end noise, the most critical area base on impedance field analysis •Clear differences between Drain- and Gatecurrent noise contributions 0.10 0.15 Gate Current Noise--Short Channel MOS µ mHz) Contribution to Gate Noise /(A • 0.18 µm channel length. Vds = 1.8 V f=5GHz 2.0x10-25 Gate Noise 2 Hydrodynamic Drift-Diffusion gate 1.5x10-25 1.0x10-25 Hot Carriers 5.0x10-26 0.0 -0.15 drain source -0.10 -0.05 0.00 0.05 Positionµm) ( 0.10 0.15 • Noise from high energy carriers has direct impact on gate noise in short channel MOS devices. Self-Heating and Scaling of Silicon Nano-Transistors Eric Pop PhD Orals 2004 Center for Integrated Systems Stanford University 2-D: Thin Body SOI (Lg = 18 nm) Mesh Engineer to ITRS Specs: LG=18 nm, tSI=4.5 nm, tOX=1 nm NSD=1e20 cm-3, NCH=1e15 cm-3 ION=1000 µA/µm, IOFF=1 µA/µm ΦGATE=4.53 eV (Mo), VDD=0.8 V E-field if W/L = 4 then Nelec ~ 2500 total! Current MONET* J·E (no Poisson) *nanoheat.stanford.edu Ultra-Thin Body SOI Scaling intrinsic baseline scaling: Intel DST/SOI Transistor (IEDM 2001) Rox = (Ri + tox/kox)/A Ri = 2x10-8 m2K/W Rco = 6.75x10-9 m2K/W other R = L/(kA) Lex ~ Lg/2 Lsd ~ Lg tsi ~ Lg/4 tsd ~ 2tsi W ~ 3Lg Aco ~ 2LgxLg extrinsic: link-up (Lex) elevated S/D (Rsd) other parameters: (Ion, Vdd, tox) from ITRS Eric Pop, Stanford PhD 2004 Thermal Modeling for SOI Drain T Rise (K) Stanford--E. Pop et al (IEDM 2003) double Lex = Lg T ↑ low power increase tsd = 3tsi T ↓ halve Rco T ↓ Gate Length Lg (nm) • “Baseline” (ITRS power scaling): tsd=2tsi, Lex=Lg/2 • “Low power” case uses proposed quadratic power scaling guidelines--reduce power, consistent with volume Proposed Power (I·V) Scaling Proposed quadratic power scaling Q = I ⋅ V = −0.17 L2g + 27.5 Lg (Lg in nm) • • • • ITRS power scaling non-uniform Quadratic scales power closer with device dimension (and volume) scaling Device temperatures near-isothermal 250 W/m power budget Vdd = 0.25 V @ Ion = 1000 A/m Stanford--E. Pop et al (IEDM 2004) Bottom-Line: GOI can achieve lower delays; improved drive current results in lower heat generation Intrinsic Delay (ps) SOI-GOI Gate-Delay Comparisons tsd = ntsi n = 1... 5 Gate Length Lg (nm) • Ge-O-I assume tGe = 3/4tSi where tSi = Lg/4 • Si more kthin reduction due to larger phonon mean free path • Ge has 2x mobility advantage, 40% lower Vdd • Delay not lowered for S/D raised beyond ~ 3 x tfilm Proposed Methodology* (RF Power MOS Devices) • Minimize fabricationcharacterization time by using device simulation • Generate table-based model from device simulation for use in circuit simulation • Resulting model: – “Black box” model – Easy enough to generate for quick evaluation of design changes *C. Ito et al, Title, SISPAD 2004 This work Device Fabrication Process Simulation Measurements Device Simulation DC, AC parameters Model Extraction Model Extraction Circuit Simulation RF Operation Metrics: Power Gain, Efficiency, Gain Compression, Intermodulation, etc. Thermal Characteristics (Simulation Results) • RF power amplifiers require temperature-dependent models • Implemented using several tables extracted at different temperatures • Characteristics for arbitrary temperatures may be linearly interpolated between tables DC Characteristics (Model vs. Measurements*) • Isothermal (pulsed) currentvoltage (IV) measurements and simulations • IV characteristics match well between model and measurement • Small discrepancy possibly due to differences between simulation doping profile and actual device profile *Infineon Device Methodology for Modeling • Compact models are becoming more abundant and complex • Implementation is becoming bottleneck and potential weak link • Cross-platform model portability is important to technology deployment • HLD Languages (Verilog/VHDL…) offer powerful solution Models Implemented in Verilog-A • • • • • • BSIM3, BSIM4.3, BSIMSOI HiSIM, Shur-RPI TFT, EKV HICUM SPICE Gummel-Poon, Diode, JFET Philips MEXTRAM, MOS 9, MOS 11 Triquent, Curtice, Parker-Skellern, Raytheon-Statz, Angelov • UCSD (Aspeck) GaAs HBT Multiple Simulators, Single Model • Compiled Verilog-A devices can be shared among diverse simulators • Same compiled object file linked to each simulator • Develop in one simulator, same results in all simulators BSIM3.cml Tiburon Design Automation Example: Adding Self-heating to BSIM3 • Adding a thermal circuit to a compact model can be a complicated project, as all of the derivatives with respect to the temperature must be provided to the simulator. • Since Verilog-A does that for the developer, the additional lines of code are minimal, typically just a few dozen, including the parameter definition. Electrical Model for Self-heating • An R-C circuit is added to the intrinsic model • The current source represents the power dissipated in the device • The voltage across the thermal resistance and thermal capacitance represents the associated temperature rise • This temperature rise is fed back to the device model Pdiss Trise Rth Tiburon Design Automation Cth Verilog-A Implementation • The code change is minimal Tiburon Design Automation DC I-V Results • The (standard) BSIM3 compared to the BSIM3 with selfheating Tiburon Design Automation Summary and Conclusions • Development of scaled-MOS compact models face many challenges: intrinsic and parasitic device effects and “other” challenges: – Intrinsic Limits--mobility, gate current, “leakage” – RF Issues--NQS, substrate effects, noise – Thermal Issues--new models, scaling laws • Methodology for compact model deployment shifting to HLL approach and compilers; “handcoded” models (for “efficiency”) inherently limit portability • “We live in interesting times (still!)” Back-up Slides • Power RF Table-based model (Root) • Silicon RTD--physics and Verilog-A model Resonant Tunneling • Higher gate tunneling in DG than SG (bulk) MOS • Simulated resonant gate tunneling current for thin layer DG SOI by using NEMO: possibly Si-based resonant tunneling diodes (RTD) € 0.0 0.5 1.0 1.5 2.0 Vg (V) € Ig (uA/um) DG (Tsi=5nm,tox=1.5nm) SG Ig (µA/µm)2 Ig (µA/µm)2 Ig (uA/um) 70 70 60 60 50 50 40 40 30 30 20 20 10 10 00 70 I=c1V[tan-1(c2V+c3)-tan-1(c2V+c4)] 70 +c5Vm+c6Vn 60 60 50 50 40 40 NEMO compact model 30 30 20 20 10 10 00 0.0 0.5 1.0 1.5 Vg (V) 2.0 Advanced Device Development • Verilog-A provides a simple way to implement new devices • Resonant tunneling diode Tiburon Design Automation Table-Based Models • Table lookup + interpolation • Separate models can easily be generated for different designs – Possible due to efficient model generation • Root Model – Consists of 5 variables over the Vg-Vd space – Igs, Ids, Qg, Qd, Idh Im{ y11} Im{ y12 } ∇QG = v gs + vds ω ω Im{ y21} Im{ y22 } ∇QD = v gs + vds ω ω ∇I Dh = Re{ y21}v gs + Re{ y22 }vds