A Model-Adaptable MOSFET Parameter Extraction System

Transcription

1 A Model-Adatable MOSFET Parameter Extraction System Masaki Kondo Hidetoshi Onodera Keikichi Tamaru Deartment of Electronics Faculty of Engineering, Kyoto University Kyoto 66-1, JAPAN Tel: Fax: Abstract--- A model-adatable arameter extraction system is develoed to catch u with raid develoment of new advanced MOSFET models. The model-adatability relies on two techniques; a model-adatable initial value estimation method and a design environment that stores and reuses extraction rocedures. The system makes it easy to develo an extraction rocedure for a new MOSFET model through the reuse of an existing rocedure for a revious model. We have resently verified that the system can accommodate major SPICE models including Level2-3 and -3. I. Introduction As the minimum feature size of MOSFETs goes into submicron and further, accurate MOSFET models for reliable circuit simulation becomes more and more imortant. Many advanced models have been develoed for keeing u with the raid rogress of rocess technologies. The roblem here is that the of a new model, esecially DC, can not be extracted easily. The arameter extraction is indisensable to characterize the MOSFETs. However, conventional arameter extraction systems are suitable only for articular models since the develoment of a arameter extraction rocedure is highly model-deendent. It is common to devise a dedicated extraction tool for each model, which is a time consuming and knowledge intensive task. In this aer, we describe a arameter extraction system that is suitable for many MOSFET DC models including newly develoed advanced ones. Two techniques are alied to build the system. The first one is a model-adatable method for accurate estimation of DC model [1]. This method contributes to eliminate the model-deendency of extraction rocedures as much as ossible. The second one is a design environment that can store and reuse extraction rocedures[2]. This system makes it ossible to reuse and modify an existing rocedure for the develoment of a new rocedure. These features reduce the time and cost for adoting a new MOSFET model. We describe extraction rocedures for major SPICE models such as -3[3, 4, ] and Meta-MOS(Level28 in HSPICE)[6], and resent the result of arameter extraction exeriments. II. Techniques for model-adatability In this section, two key techniques which are contributed to the model-adatability of the arameter extraction system are described. They are a model-adatable arameter value estimation method using a common intermediate model and a design system that stores and reuses extraction rocedures. A. Model-adatable arameter value estimation method A arameter extraction rocedure usually relies on a numerical otimization technique. It consists of initial value estimation and curve fitting with a non-linear otimization algorithm. The otimization algorithm is model-indeendent and alicable to any MOSFET models. The initial value estimation rocedure, however, is highly model-deendent. Hence, a dedicated extraction tool should be devised for each model one by one with a lot of effort and time. We have roosed an initial value estimation method which is easily alicable to many MOSFET models including newly develoed advanced models[1]. This method rovides a framework for MOSFET arameter extraction that eliminates modeldeendency as much as ossible. The rocessing flow of the initial value estimation is summarized in Fig. 1. A key idea of the method is the decomosition of the rocess into two consecutive rocesses with the use of a common and simle intermediate model. The first rocess is the extraction of intermediate model from measured I-V characteristics. This rocess is model-indeendent. The second rocess is the transformation from the intermediate model into target model. Although this rocess is modeldeendent, it has been shown that a roer intermediate model has an ability to eliminate most of the model-deendency. Given a new MOSFET model, the only task a designer has to do is to derive a rocedure for the arameter transformation which may be slightly different from rocedures for other models.

2 I-V data Extraction transform V th into target model of model. The model calculates the V th according to Intermediate model Target of model 1!&!&!&!& Target of model n Model-indeendent Transformation Systematic Fig. 1. : Processing flow for model-adatable initial value estimation. An intermediate model is shown below[1]. linear region : V 1 GS Vth 2 I DS = V DS V DS 1 1+ (V GS Vth) saturation region : not considered subthreshold region : >: I DS = const 1 ex VGS Vth N : (1) It has four key ; threshold voltage Vth, gain, mobility degradation, and subthreshold gate swing N. These are functions of V BS. They are analytically extracted by Vth = b 1 2 V DS = a b c 1 2 V DS V DS = 1 b c >: h 1 2 V DS d (lni N = DS i ) 1 (in subthreshold region) dv GS (2) where a; b and c are calculated from three sets of measured (V GS ;I DS ) values with a fixed small V DS and V BS according to the equations shown in [7]. We can exloit the deendency of the intermediate model on V BS to transform them into target model. Most models have similar structures for every intermediate model as P i (V BS )= n i X2 3 ij 1 f ij (V BS ) + fh i (V BS ): (3) j P i reresents an intermediate model arameter, and it is a function of V BS. The term ij is a target model arameter. Both f ij and fh i are functions of V BS and known model, which are derived directly from equations of the target model. As an examle, we describe a method to Vth=V FB +K 1 S V BS K 2 ( S V BS )1V DS : (4) Target model in this case are V FB ;K 1 ; and K 2. Parameter S is a known rocess arameter which reresents surface-inversion otential. The effect of is negligible under small V DS. Then, the following equations are derived for three different V BS values. >: Vth 1 =V FB +K 1 S V BS1 K 2 S V BS11 Vth 2 =V FB +K 1 S V BS2 K 2 S V BS21 Vth 3 =V FB +K 1 S V BS3 K 2 S V BS31 If Vth i is extracted under each V BS, we can solve above simultaneous equations for V FB ;K 1 ;and K 2. Many conventional arameter extraction tools also extract certain hysical model from measured I-V characteristics. A main advantage to use the intermediate model is that more can be calculated accurately and systematically than the conventional tools. () B. System that stores and reuses extraction rocedures The initial value estimation method makes arameter extraction rocedures systematic and less model-deendent. If the rocedure is described in a simle manner, we can easily develo an extraction rocedure for a new model through the reuse and modification of existing rocedures for other models. This feature imroves model-adatability of the extraction system. We can use GUIDE interactive design environment[2] to build a model-adatable arameter extraction system. The GUIDE system stores and reuses design rocedures oerated by a designer. An design rocedure is described by an Interactive Design Language(IDL) in a ste-by-ste manner, and stored in a scrit file. The stored design rocedure is ready to reuse and modify if necessary. III. Parameter extraction system In this section, our arameter extraction system is resented. Extraction rocedures for SPICE models and their re-usabilities are also discussed. A. System overview We have integrated a arameter extraction system using GUIDE design environment described in 2.2. Fig. 2 shows the structure of the arameter extraction system. IDL was originally develoed for LSI circuit and layout design[2]. We have extended the IDL for describing a arameter extraction rocedure; from I-V measurement to final curve fitting. We have added new commands for initial value estimation and for controlling measuring instruments through the GP-IB interface.

3 Extraction rocedure ( IDL in scrit file ) Systematic GUIDE Designer Command interreter Model-indeendent Interfaces Immediate result Simulator Otimizer Measuring instrument Fig. 2. : Parameter extraction system on GUIDE. The intermediate model are extracted by a secific command which solves (2). This rocess is common for all models. Then, each intermediate model arameter is systematically transformed into target model through simultaneous linear equations based on (3). The rocedure for the arameter transformation, therefore, is easily described in a ste-by-ste manner by the extended IDL. The rocedure consists of two sections; the calculation of coefficients f ij and fh i, and a command to solve the simultaneous equations. The curve fitting rocedure is also described by the IDL. In our system, a arameter extraction rocedure is comletely described by the IDL and stored in a scrit file. The stored extraction rocedure can be reused or modified easily to accommodate a new model into the system. B. Extraction rocedures for SPICE models We have develoed arameter extraction rocedures for major SPICE models. Fig. 3 shows a art of IDL descrition to extract model. The extract intmodel() command at the first line executes the extraction of intermediate model from measured data. Calculation of coefficients to transform the intermediate model arameter Vthfollows. Then, the transform tarmodel() command solves () using the calculated coefficients. Other target model are similarly transformed from,, and N. Finally, curve fitting is carried out through the otimize() command. Extraction rocedures for other SPICE models are also described using IDL. They are the same excet for arameter names and several IDL lines for the arameter transformation. Table I exlains re-usability of the transformation rocedures. We define the re-usability as a ratio of the number of target model that can be transformed by identical IDL descrition. For examle, the rocedure for the Level3 model is develoed by the reuse of the one for the Level2 model. Four out of six target model of the Level3 model can be calculated by the same transformation rocedure for the Level2 Extraction rocedure for (Level13) model extract_intmodel( vgid vth, beta, theta, n ) Transformation from Vth a11 = 1 a12 = sqrt( m1:hi - vbs1 ) a13 = vbs1 - m1:hi b1 = vth1 - m1:hi Equation at VBS = vbs1 Equation at VBS = vbs2 Equation at VBS = vbs3 transform_tarmodel( a, b m1:vfb, m1:k1, m1: k2 ) Final curve fitting otimizer( "nsol" ) otimize( "bsim1.ot" ) Measurement etc. Transformation from &B, &H, N Fig. 3. : IDL descrition for extraction. TABLE I. : Re-usability of arameter transformation rocedures from to Level2 Level3 BSIM2 Meta-MOS BSIM3 Re-usability Re-usability = reusable target model. We can also easily reuse and modify a stored rocedure for the model to develo those for new models such as BSIM2,3 and Meta-MOS(Level28 in HSPICE). IV. Exeriments First, we show the result of arameter extraction of an n-channel MOSFET with.6m channel length. The target models in this case are -3, and Meta-MOS. Table II summarizes the number of all target model whose initial values can be derived according to the method shown in 2 with the total number of otimized. The initial values of that are not derived by the method in 2 are determined randomly within their ossible ranges and sets of initial arameter values are reared for each model. Table III shows the distribution of rms-errors of I DS -V DS characteristics after otimization. We also show the average comutation time for otimization in Table IV, which is measured on SPARCstation 1. In all trials we have reached to satisfactory solutions within ractical caomutation time. The largest rms-errors of all trials is about 1.% which occurred

4 TABLE II. : The number of target and otimized target otimized TABLE III. : Distribution of final IDS-VDS rms-errors (%) on an n-channel MOSFET(L=.6m) Ave Min Max [ ma ] Drain-Source Current 3 Simulated Measured V BS = V V BS = -3.3V Drain-Source Voltage V GS = 3.3V V GS = 2.2V [ V ] Fig. 4. : Comarison between measured and simulated IDS-VDS characteristics after otimization. This is the worst case out of 2 trials. for the model. The measured and simulated I DS -V DS data of this case are comared in Figure 4. Furthermore, the G DS -V DS data of the case are comared in Figure. Although we have not considered the G DS -V DS data during the otimization, we can see that good agreement is achieved even in the worst case. Then, we have comared the results shown in Table III with those with conventional initial value estimation methods. The conventional methods considered are as follows. method(1) All of initial arameter values are determined randomly within their ossible ranges. method(2) The V th and related are calculated from measured I-V characteristics[8] and other are determined randomly. We have otimized of the -3 and Meta-MOS models from sets of initial arameter values generated by these methods. Table V shows the robability of final rmserrors being less than 1.%. The roosed method succeeds to obtain satisfactory results in all cases, while the conventional methods fail in many case. TABLE IV. : Average comutation time for otimization with SS1 cu-time 44.8s 8.9s 38.9s 9.3s TABLE V. : The robability of final rms-errors being less than 1.% roosed 1% 1% 1% 1% method(1) 2% 8% 2% 6% method(2) % 2% 8% % V. Conclusions We have resented a arameter extraction system which is suitable for many MOSFET models. The system is based on a model-adatable initial value estimation method. We have imlemented it in a GUIDE design system that stores and reuses extraction rocedures. The extraction rocedures are described by extended IDL. The stored rocedures can be reused for the extraction of a new model. The re-usability on major SPICE models is exerimentally shown. Also, the erformance of the system is demonstrated. Currently the system is targeted for the extraction of DC model. Our future work includes the develoment of extraction caability for AC. References [1] M.Kondo, H.Onodera and K.Tamaru, Model-Adatable MOSFET Parameter Extraction Method Using a Common Intermediate Model, Proc. IEEE Int l ASIC Conf., , Se [2] T.Morie, H.Onodera and K.Tamaru, A System for Analog Circuit Design that Stores and Re-uses Design Procedures, Proc. CICC 93, , May 1993.

5 1 Simulated Measured log GDS [ ma / V ] 1.1 V GS = 2.2V V GS = 3.3V.1 Drain-Source Voltage [ V ] Fig.. : Comarison between measured and simulated GDS-VDS characteristics of the worst case. [3] B.J.Sheu, D.L.Sharfetter, P.K.Ko and M.C.Jeng, BSIM:Berkeley Short-Channel IGFET Model for MOS Transistors, IEEE Jour. Solid-State Circuits, vol.sc-22, no.4,.8-66, [4] M.C.Jeng, Design and Modeling of Dee Submicrometer MOSFETs, Ph.D.Dissertation, Univ.Calif., Berkeley, [] HSPICE alication Note: BSIM3 MOS Model Level42, Meta-Software, Inc. [6] HSPICE User s Manual(H92), Meta-Software, Inc. [7] M.F.Hammer and B.Sc, First-order arameter extraction on enhancement silicon MOS transistors, IEE roceedings, vol.133, t.i,.49-4, [8] S.L.Wong and C.A.T.Salama, Imroved simulation of - and n-channel MOSFET s using an enhanced SPICE MOS3 model, IEEE Tran. Comuter-Aided Des., vol.cad-6, no.4,.86-91, 1987.