A Complete Learning-Based Semiconductor Parametric Testing and Device Modeling Ecosystem -from Probing to Simulation -

Transcription

1 Platfrm Design Autmatin, Inc. - The EDA Platfrm Cmpany A Cmplete Learning-Based Semicnductr Parametric Testing and Device Mdeling Ecsystem -frm Prbing t Simulatin - Yanfeng Li, Mia Li, Jian Ya, Rik Radjcic albert_li@platfrm-da.cm rik_radjcic@platfrm-da.cm Platfrm DA

2 A Cmplete Learning-Based Semicnductr Parametric Testing and Device Mdeling Ecsystem, frm Prbing t Simulatin Abstract Use f Artificial Intelligence methds in general, and the applicatin f specific ptimizatin techniques, neural netwrks, and learning algrithms t semicnductr parametric test and mdel generatin is described. In Parametric Test arena the gal is t imprve the quantity and quality f the data by accelerating testing speed and breaking cnventinal test hardware cnstraints. Revlutinary behavir-aware testing methds implemented at the instrument level, DUT level and prductin test level are utlined. Use f learning algrithms t autmate mdel generatin is als utlined. The paper presents key cncepts and summarizes sme specific results. Platfrm DA

3 Outline Cmpany Intrductin Artificial Intelligence Overview AI and Semicnductr Technlgy Fr Test and Measurement Fr Mdel Extractin Case Studies fr Behavir-Aware Test Methds At instrument level: reduce settling time At DUT level: curve recvery technique At prductin test levels: reduce test samples Case Study fr Autmate Mdel Generatin Bi-directin Netwrk (BDN) Cnclusin Platfrm DA

4 Outline Platfrm DA

5 Platfrm Design Autmatin (PDA): Where We Fit Bridging the space between Si technlgy and Design Intersect: Parametric Measurements, SPICE mdels and PDK s Prgressively mre int Device Physics & Material Science, etc,, Prcess Dmain (fundries, OSATS, Equipment Vendrs, Material Suppliers ) ~$300B Semicnductr Industry Mdel & Simulatin Design Dmain (Design Huses, EDA, IP vendrs, etc ) Prgressively Mre int System Architecture & High Abstractin PDA: Unique Integratin f All the Critical Elements that cnnect Test Chip Prcess Technlgy with Prduct Design Test Test Slutin Slutin Our Space Data Data Mdel Mdel PDK PDK EDA + IP Platfrm DA

6 Cmpany Intrductin PDA : Big Fish in a Small Pnd Funded in July 2012 Frmer Accelicn (acquired by Agilent in 2012) Experienced team with a slid track recrd frm Cadence, PDF Slutins, Keysight, Qualcmm... Beijing Office Offering: Measure-Design-Prcess Integratin Integrated Measurement, Mdeling and Design Slutins 1-Stp design infrastructure services Beijing Lab Cre Cmpetence Device characterizatin, mdeling and PDK Artificial Intelligence (AI) algrithms Value Prpsitin: Efficiency & Time & Value Fastest Test = Mst DATA Fastest Prductin Parametric Tests = Reduce TTM Algrithms t break hardware cnstraints = Mst Capable Shanghai Office TW Hsinchu Office Platfrm DA

7 Cmpany Intrductin PDA : Big Fish in a Small Pnd Funded in July 2012 Frmer Accelicn (acquired by Agilent in 2012) Experienced team with a slid track recrd frm Cadence, PDF Slutins, Keysight, Qualcmm... Beijing Office Offering: Measure-Design-Prcess Integratin Integrated Measurement, Mdeling and Design Slutins 1-Stp design infrastructure services Beijing Lab Cre Cmpetence Device characterizatin, mdeling and PDK Artificial Intelligence (AI) algrithms Value Prpsitin: Efficiency & Time & Value Fastest Test = Mst DATA Fastest Prductin Parametric Tests = Reduce TTM Algrithms t break hardware cnstraints = Mst Capable Shanghai Office TW Hsinchu Office Platfrm DA

8 Prduct and Service Prtflis Test & Characterizatin Device Mdeling PDK Services Prducts FS360 & FS380 NC300 & NC300L SEK MeQLab FastLab Artificial Intelligence PQLab Characterizatin Services Mdeling Service PDK Generatin Service + Cell Lib & Cmpiler Service Platfrm DA

9 Outline Cmpany Intrductin Artificial Intelligence Overview AI and Semicnductr Technlgy Fr Test and Measurement Fr Mdel Extractin Case Studies fr Behavir-Aware Test Methds At instrument level: reduce settling time At DUT level: curve recvery technique At prductin test levels: reduce test samples Case Study fr Autmate Mdel Generatin Bi-directin Netwrk (BDN) Cnclusin Platfrm DA

10 Artificial Intelligence : Overview (1) A lt f buzz abut Artificial Intelligence AI: any device that perceives its envirnment and takes actins that maximize its chance f success at sme gal Machine Learning: algrithms that avid fllwing static prgram instructins and can learn by building a mdel frm sample input data and make data-driven decisins Deep Learning: based n learning data representatins, vs task-specific algrithms. Learning can be supervised, partially supervised r unsupervised Deep Neural Netwrks : (DNN) is an artificial neural netwrk (ANN) with multiple hidden layers between the input and utput layers and can mdel cmplex nn-linear relatinships Recurrent Neural Netwrks: (RNN) is a class f artificial neural netwrk (ANN) where cnnectins between units frm a directed cycle. This allws it t exhibit dynamic tempral behavir. Lts f Data + Cheap Cmpute Pwer Data Driven Training Sets (Supervised and Unsupervised) Prcessing Pwer t Derive Mdels Applicatins Cmputer Visin & Vice Recgnitin Autnmus Driving Oh and, Spam Filter, Credit Card Prtectin Platfrm DA

11 Artificial Intelligence : Overview (1) A lt f buzz abut Artificial Intelligence AI: any device that perceives its envirnment and takes actins that maximize its chance f success at sme gal Machine Learning: algrithms that avid fllwing static prgram instructins and can learn by building a mdel frm sample input data and make data-driven decisins Deep Learning: based n learning data representatins, vs task-specific algrithms. Learning can be supervised, partially supervised r unsupervised Deep Neural Netwrks : (DNN) is an artificial neural netwrk (ANN) with multiple hidden layers between the input and utput layers and can mdel cmplex nn-linear relatinships Recurrent Neural Netwrks: (RNN) is a class f artificial neural netwrk (ANN) where cnnectins between units frm a directed cycle. This allws it t exhibit dynamic tempral behavir. Been arund since 1956 What is New and Why Nw Lts f Data + Cheap Cmpute Pwer Data Driven Training Sets (Supervised and Unsupervised) Prcessing Pwer t Derive Mdels Applicatins Cmputer Visin & Vice Recgnitin Autnmus Driving Oh and, Spam Filter, Credit Card Prtectin This is New-ish And driving the buzz This is Our Wrld same ld... Platfrm DA

12 Artificial Intelligence : Overview (2) Spectrum f Algrithm Cmplexity & Data Requirements Data Behavir Text Recgnitin Chess Vice Recgnitin GO Autnmus Driving Facial Recgnitin Self Driving Trains Algrithm Cmplexity/Difficulty Platfrm DA

13 Artificial Intelligence : Overview (2) Spectrum f Algrithm Cmplexity & Data Requirements Data Behavir Semicnductr Parametric Data Vice Recgnitin GO Autnmus Driving Facial Recgnitin Self Driving Trains Algrithm Cmplexity/Difficulty Our Wrld is On the Lw End f the Spectrum Mstly cnstrained and predictable behavir Lts f histry and physics and many data pints And n ne dies if errr rate is > % Platfrm DA

14 Outline Cmpany Intrductin Artificial Intelligence Overview AI and Semicnductr Technlgy Fr Test and Measurement Fr Mdel Extractin Case Studies fr Behavir-Aware Test Methds At instrument level: reduce settling time At DUT level: curve recvery technique At prductin test levels: reduce test samples Case Study fr Autmate Mdel Generatin Bi-directin Netwrk (BDN) Cnclusin Platfrm DA

15 AI and Semicnductr Technlgy Technlgy Trends Shrinking Margins & Grwing Variability Prliferatin in Technlgy Flavrs What Des the Industry Need Mre Data Fr prcess cntrl Fr Design Targeting Less Test Time Cnstrain Csts and Enhance Prductivity Opprtunity fr Applicatin f AI Enhance Accuracy = Reslve Variability Accelerate Test = Reduce Test Cst Mre DFM = Mdels Platfrm DA

16 AI and Semicnductr Technlgy Technlgy Trends Shrinking Margins & Grwing Variability Prliferatin in Technlgy Flavrs What Des the Industry Need Mre Data Fr prcess cntrl Fr Design Targeting Less Test Time Cnstrain Csts and Enhance Prductivity Opprtunity fr Applicatin f AI Enhance Accuracy = Reslve Variability Accelerate Test = Reduce Test Cst Mre DFM = Mdels Mtherhd & Apple Pie But Still True Platfrm DA

17 Data Standard flw f machine learning Preprcessing Feature Representatin Feature extractn. Feature selectin Learning algrithm Inference, Predictin, Recgnitin Mst critical fr accuracy Accunt fr mst f the cmputatin fr testing Mst time-cnsuming in develpment cycle Often hand-crafted in practice Mst Effrt in Machine Learning Slide Curtesy: Andrew Ng, Kai Yu Hw d we apply this standard flw specifically fr IC Industry? IIP3(dBm) NF(dB) 6 Samples Paret Set Paret Frnt Gain(dB) Platfrm DA

18 Step 1: Data Data Preprcessing Feature extractn. Feature selectin Inference, Predictin, Recgnitin Traditinal New Era Measure Measure Mre Stre Stre & Mine Measurement des nt g Away (there is n magic) In fact : need mre f it => Must be faster & cheaper New effects, New variability, Less margin Must Leverage the Data t Enhance Measurements Data mining shaping the data gathering Platfrm DA

19 Step 2: Feature Representatin Data Preprcessing Feature extract. Feature selectin Inference, Predictin, Recgnitin Data Simplificatin Cmputer visin preprcessing IC industry Data Cmpressin Techniques Infrmatin Cmpressin Principal Cmpnent Analysis (PCA): successfully used fr crner mdels Fast Furier Transfrmatin (FFT): cmmn in signal prcessing Slide Curtesy: Andrew Ng Standard Practices used in Semicnductr Technlgy are Analgus t Prcedures Used fr high end AI Techniques such as PCA, FFT, Optimizatin Vs preprcessing fr cmputer visn Platfrm DA

20 Step 2: Feature Representatin Data Preprcessing Feature extract. Feature selectin Inference, Predictin, Recgnitin Deep Learning Methds Aut Feature Extractin IC Industry Mdeling Practices Optimizatin Unsupervised learning between measurement and mdeling t reach the ptimized SPICE mdel Analgus Algrithms and Structures may be Applied Artificial Neural Netwrk (ANN) t imitate brain Cnvlutinal Neural Netwrk (CNN) Recurrent Neural Netwrks (RNN) Platfrm DA

21 Step 3: Inference & Predictin Data Preprcessing Feature extract Feature selectin Inference, Predictin, Recgnitin AI Methdlgies Late 80 s Neural Netwrks;Bsting;Supprt Vectr Machines;Maximum Entrpy Since 2000 learning with structures Kernel Learning; Transfer Learning; Manifld Learning; Sparse Learning Matrix Factrizatin; Structured Input-Output Predictin; Mainstream Applicatins Semicnductr Applicatins Faster Measurements Autmatic Mdeling Autmatic Design? Platfrm DA

22 Step 3 : in Semicnductr Technlgy Traditinal Versin f Predictin & Inference => Mdeling & Simulatin Use Fitting & Interplatin Techniques t Optimize Mdel Leverage Physical Mdels t Define/Cnstrain Curve Shape AI Opprtunity? : Relativity Learning define relativity functin (relatinship) between each pint Analgus t techniques like in mdels used t predict mtin tracking in graphic prcessing Opprtunity t Use AI and D Things Differently? Platfrm DA

23 Learning Technique vs. Traditinal Fit Blue:Relativity Learning Red:Spline Blue:Relativity Learning Red:Spline Resistr Dide 一阶导 Blue:Relativity Learning Red:Spline Blue:Relativity Learning Red:Spline BJT MOS Platfrm DA

24 Outline Cmpany Intrductin Artificial Intelligence Overview AI and Semicnductr Technlgy Fr Test and Measurement Fr Mdel Extractin Case Studies fr Behavir-Aware Test Methds At instrument level: reduce settling time At DUT level: curve recvery At prductin test levels: reduce test samples Case Study fr Autmate Mdel Generatin Bi-directin Netwrk (BDN) Cnclusin Platfrm DA

25 Case Studies: At instrument level Reduce Settling Time AI fr Adaptive Starting Pint Selectin Result : accelerated test time Super Fast Starting Pint Fast Starting Pint Traditinal Starting Pint Real Signal Machine Learning (ML) predictin Time Series Analysis (TSA) predictin Platfrm DA

26 Case Studies: At DUT level Principle: Leverage Expected DUT Behavir t Imprve Test DUT Characteristics are Knwn and can be Anticipated e.g. I-V, C-V, 1/f vs L,W,T Histrical and Relative data e.g. same device / different bias pint e.g. same TEG / different device size e.g. same wafer / different TEG e.g. same structure / different wafer etc Verified mdel e.g. BSIM I V, 1/f flicker nise.. Curve Recvery algrithms e.g. Relativity Learning Platfrm DA

27 Case Studies: At DUT level Many Opprtunities t Accelerate Test w/ Lss f Accuracy simple => cmplex e.g. Pre-Set the SMU Range since yu knw what t expect e.g. Sparsify Test Step Size Sample & Curve Recvery Relativity Learning algrithms Can still extract all desired derivatives e.g. De-embed signal frm nise By applying suitable dmain transfrmatins e.g. device nise vs test system nise Platfrm DA

28 Case Studies: At Prductin Test level Bad Case Detectin e.g. Bad Prbe Cntact e.g. Bad Device, TEG r Wafer 1. Use Machine Learning t Define Envelpe f Expected Behavir Inc. all SPC allwed Variability 2. Cmpare Measured Data t Expected Values 3. If Data Outside Envelpe => Abrt Simple But Effective (D nt Cllect Bad Data that is Dumped Later) Esp. Valuable fr Lng Autmated Test Rutines e.g. Overnight WAT Test Platfrm DA

29 Outline Cmpany Intrductin Artificial Intelligence Overview AI and Semicnductr Technlgy Fr Test and Measurement Fr Mdel Extractin Case Studies fr Behavir-Aware Test Methds At instrument level: reduce settling time At DUT level: curve recvery technique At prductin test levels: reduce test samples Case Study fr Autmate Mdel Generatin Bi-directin Netwrk (BDN) Cnclusin Platfrm DA

30 Mdel Generatin Practices State f the Art BSIM Mdel: Fundamental principles haven t changed in 20 years. Fitting targets (In, Vth, Gm, etc.) haven t changed fr 20 years Have a lt f expertise & data and we accept sme fitting errrs Traditinal Appraches Curve fitting (e.g., linear regressin) Require physical input Pre-defined curve shape Curve interplatin (e.g., spline) Nise in data Parameter Tweaking Changing Criteria Runge's phenmena Cannt derive Cnfidence Level Platfrm DA

31 Mdel Generatin Practices State f the Art BSIM Mdel: Fundamental principles haven t changed in 20 years. Fitting targets (In, Vth, Gm, etc.) haven t changed fr 20 years Have a lt f expertise & data and we accept sme fitting errrs Ideal Applicatin fr Bidirectinal Recurrent Neural Netwrks (BRNN) d nt require input data t be fixed. future input infrmatin is reachable frm the current state. Parameter Tweaking Changing Criteria Platfrm DA

32 Mdel Generatin vs Mdel Selectin Which is Best? Depends n Design r The real art is in understanding the tradeffs and selecting the right slutin fr a given applicatin nt in tweaking parameters Platfrm DA

33 Case Study:BRNN fr Mdeling Apply BRNN t Prduce a set f Prpsed Mdels 1. Use Machine Learning t prpse varius BSIM Parameters 2. Device Engineer Selects the Best Match fr his Target 3. Supervised training per nde and per device type => imprve ver time Prduce multiple answers based n Machine Learning N iterative ptimizatin and parameter tuning Prpsed Candidates d Nt have t be exact (e.g. SIRI vs Search Engine) 3 Minutes/Rund 40x23=920 specs 40x17=680 parameters QA cnstraints Mdeling engineers Train the system and Select the Best Optin Judges rather than spending time n tweaking parameters AI replaces labr nt expertise e.g. knw yur design needs & use expertise t select the best slutin Platfrm DA

34 Outline Cmpany Intrductin Artificial Intelligence Overview AI and Semicnductr Technlgy Fr Test and Measurement Fr Mdel Extractin Case Studies fr Behavir-Aware Test Methds At instrument level: reduce settling time At DUT level: curve recvery technique At prductin test levels: reduce test samples Case Study fr Autmate Mdel Generatin Bi-directin Netwrk (BDN) Cnclusin Platfrm DA

35 Practicing DARK Arts Data Algrithms Risk Knwledge Accumulated millins f curves fr different fundry/prcesses We are ptimizatin experts with daily exercises with different neural netwrks We Shuld expect and Manage the errr in simulatin r measurement Mdeling backgrund gives us the best knwledge in device/ic behavirs AI Lks Very Prmising => We shuld Embrace It Nt Mysterius Magic Semicnductr Technlgy is Characterized by Cntained Scpe This is Imprtant & Makes Applicatin f AI Easier Leave ut the Fancy Wrds and Lets Get Practical with AI! Platfrm DA