Slide 1 Developing Reliable Device Simulation Models using ADS ADS User Meeting Böblingen, 14.05.2009 ADS User Meeting Böblingen May 2009 Dr. Thomas Gneiting AdMOS GmbH Frickenhausen [email protected] Dr. Franz Sischka Agilent Technologies Böblingen [email protected] 1 (C) Bildarchiv der Stadt Böblingen Slide 2 Developing reliable device simulation models using ADS AGENDA - The difference between selecting a model and extracting model parameters. - Trusted, verified measurements. - contact resistance - self-heating effects - linear S-parameter - de-embedding - Example of a device modeling sequence 2 Slide 3 Which model to select ? Semiconductor Device Models (Nonlinear Models): - BJT Models: HICUM 2, HICUM 0, MEXTRAM, VBIC, Gummel-Poon, ST BJT, Agilent EEBJT2 - Diode Models: PN diodes, Philips JUNCAP, Agilent Root, PIN - MESFET/HEMT Models: TOM3, Angelov, Agilent EEFET3 and EEHEMT1, Agilent Root MESFET/HEMT (HPFET), TOM scaleable, Curtice quadratic and cubic, Advanced Curtice, Statz et al. (Raytheon), Materka, Modified Materka (Rizzoli), Tajima - MOSFET Models: BSIM3, BSIM4, BSIM3SOI, BSIM4SOI, PSP, HiSIM2, HiSIM-HV, Philips MOS Model 11, Philips MOS Model 30, Philips MOS, Model 9, Agilent Root MOSFET, Agilent EEMOS1, BSIM1, BSIM2, MOS Level 1,2,3 3 There are many models available, even for the same device ! Slide 4 Which model is better ? Model A ? Model B Only reliable measurements provide good simulation models 4 You may think of using rather model_A than model_B. However, as we will show with the following slide set, the main topic is to perform reliable, verified measurements. This is the very base of any good model implemented in a design kit of a manufacturer. Of course, some models feature equations to better fit a certain effect (thermal self-heating, non-linear large-signal RF behavior). But also here applies the statement of above: if the measurements are not available, the ‘better’ model does not help. And if the measurements are not ‘qualified and verified’, the ‘better’ mdoel does not help too. Slide 5 iD What is a model ? RS RS = ∂(v D − v D int ) ∂iD Ohmic Losses DC space charge vD diffusion capacitance vDint Model equations v D int iD = IS ⋅ e N* vt − 1 DC vt = 8.6171E - 5 ⋅ (TEMP + 273.15 ) Cs = C JO v D int 1 − VJ Junction Cap M CD = TT ⋅ ∂iD 1 = TT ⋅ ⋅ iD ∂v D int N ⋅ vt Diff.Cap (Transit Time) Is ... Model Parameter 5 A set of equations describing the principal electrical behavior of a device represents what is commonly called a device model. NOTE: A ‘SPICE simulation model’ is a name for a model available in UCB Spice (University of California, Berkeley), and thus typically available in all of today’s simulators. For a real device, only the model parameters like e.g. IS, NF are adjusted to get a simulation results close to the real behavior. Additionally, the instance parameters (e.g. L, W) describe the geometry and structure of a device. If no model parameters or instance parameters are specified, the simulator takes default parameter values. They shall prevent the simulator from numeric errors (e.g. divide by zero) and represent an ideal device (the specific model effect is ‘switched-off’). Caution: Default parameter are in most cases DC bias and frequency independent!! Slide 6 What is model parameter extraction ? Plot diode_basics/DC/forward/ia_va 1E-1 1E-2 vD N* vt iD = IS ⋅ e − 1 1E-3 1E-4 1E-5 1E-6 Fitting the model equation to the measured data by varying the model parameters. ia.m [LOG] 1E-7 1E-8 1 decade 1E-9 2.3*N*vt 1E-10 1E-11 1E-12 1E-13 1E-14 1E-15 0.0 Is 0.2 0.4 0.6 va [E+0] 0.8 1.0 NOTE: The model equation specifies the measurement setup (what to measure)! 6 Slide 7 Model Parameters and Instance Parameters IMPORTANT NOTE: when not specified, the instance and model parameter values used in the simulation will be the ‘default values‘, i.e. the effect is switched off (CDS=0F, RTH=0Ω Ω , VAF=infinite), or the simulation is standardized (L=1). SA, SB Device Instance (L, W, geometry, lithography, size of the production masks) L W Device Model (typically process-dependent Model Parameters: VTH, U0, CDS etc.) 7 Slide 8 Developing reliable device simulation models using ADS AGENDA - The difference between selecting a model and extracting model parameters. - Trusted, verified measurements. - contact resistance - self-heating effects - linear S-parameter - de-embedding - Example of a device modeling sequence 8 Slide 9 Contact resistance of on-wafer measurements IC-CAP/ADS (C) Cascade Microtech 9 One of the most important preconsiderations regarding accurate device modeling is to characterize the contact resistance of the probes to the contact pads on the wafer. This is especially important for silicon processes, where the on-wafer contacts are made from aluminum (III/IV processes use typically gold). When not considering these contact resistances, their effect will be included in the ohmic parameters of the model, (they will become too big). This will badly affect any chip design. Slide 10 The effect of contact resistance including Rcontact id.M id id without simulatedRcontact without Rc ididwith simulated Rc [E-3] [1E-3] Plot ADS_simul_types/DC_mdlg/idvd/idvsvd 50 Behavior of the final model in design kit = real device behavior ! (without contact resistance) 40 30 20 10 Drain 0 0.0 0.5 1.0 1.5 Rcontact2 2.0 vd [E+0] IMPORTANT NOTE: Including the contact resistance is mandatory during the modeling process Measurement and Simulation including contact resistance Gate MOSFET1 Rcontact1 10 NOTE: do not confuse these measurement-related contact resistances with the device‘s inner, ohmic model parameters (e.g. RD, RS etc.) Slide 11 Self-heating with DC measurements Silicon Bipolar Transistor on wafer Ic (mA) Measurements at various pulse widths infinite 4 µsec 2 µsec 1 µsec Only a few models can handle self-heating: MOS: BSIMSOI3, BSIMSOI4, HiSIM-HV BJT: VBIC, HiCUM MESFET/HEMT: Angelov, ... Model Parameters to model the self-heating: RTH, CTH pulse period: 1msec VCE (V) Hint: 50mA device current: beginning of self-heating 11 As a general observation, currents above ~25mA will cause self-heating effects with transistors. Rule of thumb: 25mA with typically 2V equals already 50mΩ for that tiny transistor on the wafer! Self-heating can only be avoided when applying pulsed DC measurements. However, the max. duration for such a DC pulse is 100ns, followed by typically 1ms wait time. If you have to live with self-heating, make sure to have repeatable measurements for the DC Settings (currents for the transfer and for the output characteristics must be identical for identical bias conditions !!). Furthermore, the self-heating during the (very often fast measured (!)) DC measurements must be the same as with the (slowly performed) S-parameter measurements !! -> measure your DC curves as slowly as later the DC-biased S-parameters. Slide 12 S-parameter measurements Transistor and Diode S-Parameter measurements are performed including a DC bias. Due to the nonlinear transistor behavior, a too big AC amplitude results in a distorted output signal. These distortions will shift the DC operating point ! iD DC oper.point shifted by too big an RF signal DC oper.point w/o RF signal vGS 12 The energy for the harmonics of a non-linear behavior is provided - by the DC bias - by the applied RF signal. In any case, when harmonics occur, the DC bias is affected. The currents provided by the DC power supply will become a function of the distortion of the RF signals. Slide 13 Find Out The Max. Applicable RF Signal to obtain linear S-parameters Simultaneous measurement of DC behavior with overlaid S-parameter signal If the RF signal is small enough, it does not disturb the DC biasing of the S-parameter measurements ! -20dBm: RF signal disturbs DC traces -30dBm: correct RF signal level for this transistor 13 From the plot above, we can determine 2 important things for obtaining linear S-parameters: the max. applicable RF signal level for the S-parameter measurements for a given DC operating point, or, inversely, the minimum applicable DC bias for a given RF signal. Slide 14 NWA Calibration, applpying the identified RF power After the standard calibration, the calibration plane is at the end of the GSG probes. Probes and calibration substrate must match ! ISS calibration substrate Cal.Plane 14 Slide 15 S= B NWA Cal Plane G S= B D Unfortunately, our transistor is not directly at the probe contact location … 15 In the above slide, in the magnified view, is the inner transistor, which we want to characterize. All the rest has to be de-embedded, i.e. to be stripped-off. In other words, the NWA calibration plane, obtained by the NWA calibration so far, has to be shifted from the GSG probe contact location down to the beginning of the device. Slide 16 De-embedding: Shifting NWA Calibration Plane vs. DUT eliminate parasitic effects Reference plane of NWA por t 1 Ground Ground Signal Signal Ground Ground Reference plane of NWA por t 1 through deem bedding Source + Substrat Gate Drain Source + Substrat Ground Ground The simulation model shall only describe the behavior of the DUT without any influence of the surrounding pads -> The influence of the parasitic components between the NWA calibration plane and the transistor is eliminated by de-embedding. Signal Signal Ground Ground Reference plane of cir cuit library elem ent (C) www.ihp-microelectronics.com 16 Slide 17 De-embedding: Measurement Data after de-embedding measured data of device and pad 17 The two diagrams demonstrate the effect of the de-embedding. The left one shows the input and output reflection S11 and S22 before and after the de-embedding. Looking at S22, the difference in both, phase and magnitude can be seen very clearly. A similar behavior can be observed in the transmission behavior. The original measured data of the forward transmission S21 clearly has an increased phase shift and a decreased magnitude with increasing frequency, compared to the de-embedded data. As a summary, it is absolutely necessary to perform a correct de-embedding to get the real transistor data as a base for accurate modeling. Slide 18 De-embedding: Principal Teststructure Layouts OPEN SHORT OPEN and SHORT are a must for de-embedding THRU The THRU device is for verifying the de-embedding quality: -> After de-embedding from OPEN and SHORT, the THRU S-parameters must represent a short transmission line. These facts should be widely known and are explained in many publications. However, we still see many cases, where de-embedding cannot be done correctly due to bad de-embedding structures !! (C) www.ihp-microelectronics.com 18 The pre requisite for a correct de-embedding is that certain test structures are available on a wafer together with the device under test (DUT) itself. Depending on the selected de-embedding method, an OPEN, a SHORT and a THROUGH dummy pad structure must be available and must be measured. The principle layout of these structures are given above. RECOMMENDATION: the on-wafer THROUGH is an ideal, golden de-embedding verification device. Compared to a diode or transistor, the de-embedded THRU is - DC bias independent - not suffering from self-heating - not susceptible to large RF signals etc. In other words, when the THRU, de-embedded from the OPEN or the OPEN/SHORT represents a short strip line, with its characteristic appearance in both, the Smith Chart (S11, S22) and the Polar Diagram (S21, S12), it can be assumed that when replacing that little piece of aluminum by the diode or the transistor, that the de-embedding will work OK also for these very devices. Slide 19 Wrap-up: De-embedding, the step after NWA calibration The measured S-parameters of the DUT describe the behavior of the device plus the pads measuredDUT G G S S G G The DUT to-be-modeled is without the pads de-embedded DUT the DUT itself (!) Layout screen shots: (C) www.ihp-microelectronics.com 19 Slide 20 Device Modeling: The extrapolated starting points of the S-par (0Hz) are determined by the DC fitting !! If this starting point check fails, verify: -> too much RF signal -> self-heating -> voltage drop in S-par testset -> accurate DC contact resistance -> DC modeling was performed at different bias range than the S-par modeling ! measured Spar fmin=100MHz DC data converted to S-parameter, f=0Hz 20 Since S-parameter modeling means (black-and-white) characterizing ‘the speed’ from going from the DC biasing to freq->infinite, the capacitors and the transit time are the only parameters to adjust. And these capacitors and the transit time, although bias dependent, describe ‘just’ the phase(freq) and magnitudes(freq) of the S-parameters with increasing frequency. -> The extrapolated 0Hz starting points are 100% determined by the DC fitting ! Slide 21 Developing reliable device simulation models using ADS AGENDA - The difference between selecting a model and extracting model parameters. - Trusted, verified measurements. - contact resistance - self-heating effects - linear S-parameter - de-embedding - Example of a device modeling sequence 21 Slide 22 Typical Device Modeling Flow Test model - in dedicated simulators - for specified operating conditions - outside specifications (model robustness) Perform all necessary measurements Documentation Quality Assurance Generate a final report with a comparison measurement / simulation Statistical Modeling Parameter Extraction Measurements for Modeling Parametric Tests topics discussed in this presentation Extract model parameters Adjust model to process targets. Adding corner cases and statistical distributions - Golden Die selection - Mismatch & statistical analysis - Process stability 22 Slide 23 IC-CAP: Dedicated Software to Support Modeling • Parameter extraction for state-of-the-art CMOS processes (e.g. 65nm) is a highly complex task. • To fulfill all requirements with respect to accuracy and statistical behavior dedicated software is a must. • The screenshot shows the PSP Modeling Package of IC-CAP 23 Slide 24 Examples of Parameter Extractions (RF CMOS) • The following slides show different examples how parameters are adjusted. • A typical RF CMOS process with a minimum feature length of 130nm was taken as a test vehicle. • We demonstrate from which characteristics certain model parameters are extracted. • In the same simulation setup, the influence of the extracted parameter to other regions of operation are visualized. • For this purpose, the following measurements/simulations are taken into account: – DC, S-parameter, Harmonic Balance, Minimum Noise Figure 24 Slide 25 Influence of Contact Resistance on DC, S21, ... Simplified MOS model Rd Gate Rcontact1 Drain Rgate Cgd gm Rearly Rcontact2 Cgs Rs 25 Slide 26 Influence of Contact Resistance on DC, S21, ... Rcontact: 3Ω (=correct value) 10Ω Extraction: From DC Measurements of an on-wafer SHORT dummy 26 Slide 27 Influence of Gate Resistance on S11, NFMIN Simplified MOS model Rd Gate Rcontact1 Drain Rgate Cgd gm Rearly Rcontact2 Cgs Rs 27 Slide 28 Influence of Gate Resistance on S11, NFMIN Rshg: 4.45Ω/ (=correct value) 10Ω/ Extraction: From S11, Y11 NOTE: Rshg is the Gate Sheet Resistor. 28 Slide 29 Influence of DC Parameter to Id-Vd, gds, S21, S22, Pout, NFmin Simplified MOS model Rd Gate Rcontact1 Drain Rgate Cgd gm Rearly Rcontact2 Cgs Rs PCLM 0.3 .. 1.3 29 Slide 30 Influence of DC Parameter to Id-Vd, gds, S21, S22, Pout, NFmin PCLM: 1.3 (=correct value) 0.1 Extraction from: Id-Vd, gds NOTE: PCLM is a DC related parameter (Early voltage). It affects all regions of operations ! 30 Slide 31 Influence of Capacitance Parameter Simplified MOS model Rd Gate Rcontact1 Drain Rgate Cgd gm Rearly Rcontact2 Cgs Rs 31 Slide 32 Influence of Capacitance Parameter CGDL: 493pF/m (=correct value) 5nF/m Extraction from: S12 and CV 32 Slide 33 Modeling Services CMOS Technology Design libraries for CMOS and other technologies: - Advanced Processes down to 45nm - RF CMOS - High voltage devices - Silicon on Insulator - Packaged devices Passive Devices Simulation models for: - On-chip passives (inductors, ...) - PCB elements for high speed circuits - Connectors - IC packages The principal services and products of Adanced Modeling Solutions are in the area of device modeling and connector design (signal integrity). Slide 34 AdMOS Company Information • • AdMOS GmbH Advanced Modeling Solutions In den Gernaeckern 8 D-72636 Frickenhausen/Germany Phone: +49 (7025) 911698-0 Fax: +49 (7025) 911698-99 • email: [email protected] http://www.admos.de • AdMOS was founded in 1997 by Dr. Thomas Gneiting. AdMOS is focused on: – Software development of tools for model parameter extraction of CMOS and other devices. – Modeling and simulation service for complex devices and systems. – Engineering service for design and test of RF and high speed components. Actually, we employ 5 highly qualified engineers for our modeling service and software development activities. Due to our ongoing expansion, we moved to a new office building in December 2007. 34 Advanced Modeling Solutions is a young startup company, which focuses on the aspects of device modeling for circuit simulation. It was founded in 1997 by Dr. Thomas Gneiting. Advanced Modeling Solutions is located pretty close to Stuttgart, which is well known for its famous car manufacturers and a lot of high tech industry. Slide 35 Developing reliable device simulation models using ADS Wrap-Up - The difference between selecting a model and extracting model parameters. - Trusted, verified measurements. - contact resistance - self-heating effects - linear S-parameter - de-embedding - Example of a device modeling sequence 35