Full-chip simulation in smart-power IC’s ROBUSPIC Workshop A. Baguenier, ESSDERC ’06 – Montreux,Switzerland Friday 22nd of September ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 1 Outline On The Way To Trillion Dollar Markets Electrical Third generation simulation technology Reliability simulation flow Behavioural languages Classical errors Unused solver features Conclusion ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 2 On The Way To $Trillion Dollar Markets Margin pressures. Increased competition. Globalization Time To Market Productivity Design Design Start Start Trends Trends (%) (%) IC IC Dev. Dev. Costs Costs ($M) ($M) 100 25 25 20 20 15 15 10 10 55 .18µ .13µ 90nm 65nm ESSDERC ’06, Montreux Complexity Design Design Team Team Composition Composition 450 450 ROW 60 Japan Europe 20 North America 350 350 250 250 HW SW 150 150 50 50 ‘00 ‘02 ‘04 ‘06 ‘08 ROBUSPIC Workshop ‘10 .18u 65nm .18u .13u .13u 90nm 90nm 65nm A. Baguenier – Slide 3 On The Way To Trillion Dollar Markets It takes a global design chain Device Mfg. Design Chain Spec and Verification Plan PCB Design SW Design PCB Mfg: Silicon/SW/PCB Integration & Testing System Design & Verification Chip Spec Architecture Design Implementation IC/Package Mfg AMS or Digital IC Design ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 4 Verification Challenges in Nanometer Designs Design Issues Verification Need Larger designs with more analog & mixed-signal content Capacity & performance with minimal accuracy trade-off for analog Inter-connect nanometer effects, capacitive cross-coupling & inductance Accurate simulation of crosstalk noise and other signal integrity problems Critical timing delays due to RC and cross-coupling capacitance Seamless post-layout simulation flow with throughput to handle larger data Degradation in timing and failures due to IR drop at low supplies Accurate simulation of IR drop effects with capacity to handle large designs Mixed-signal designs with tighter interaction between analog & digital Ability to accurately simulate digital & analog components as a system ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 5 Smart Power IC => Many different voltages 0 to 1Kv 0 to 1Kv 0 to 1Kv 0 to 12v Different tolerances for each blocs. Local option setting Multi- rate simulation in the control Unit ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 6 Simulator Generations FirstGeneration Generation First Spectre Spectre SPICE SPICE SecondGeneration Generation Second ThirdGeneration Generation Third FastSPICE FastSPICE Ex:UltraSim UltraSim Ex: Technology Technology Technology Technology Technology Technology CompactModels Models Compact SparseMatrix MatrixSolver Solver Sparse UniformTime TimeStep Step Uniform NewtonRaphson Raphson Newton SimplifiedModels Models Simplified MatrixPartitioning Partitioning Matrix EventDriven Driven Event MultiRate Rate Multi RCReduction Reduction RC Hierarchy Hierarchy Isomorphism Isomorphism AdaptivePartitioning Partitioning Adaptive Applications Applications Applications Applications Applications Applications BlockDesign Design Block Analog/MS/RF Analog/MS/RF Prelayout Prelayout Tran,AC, AC,Noise Noise Tran, 50Kdevice devicecapacity capacity 50K ESSDERC ’06, Montreux LargeBlock BlockDesign Design Large MS/Digital/Memory MS/Digital/Memory Pre-/Postlayout Pre-/Postlayout Tran Tran 10Mdevice devicecapacity capacity 10M ROBUSPIC Workshop FullChip Chip Full MS/Digital/Mem/RF MS/Digital/Mem/RF Pre-/Postlayout Pre-/Postlayout Tran Tran 1B+device devicecapacity capacity 1B+ A. Baguenier – Slide 7 Event Driven Simulation / Partitioning FastSPICE Conventional SPICE A X1 X2 One Partition One Matrix . . . . . G1 . . . . . . G2 . . . G3 . . . . . . . G4 . . . . . . . G5 . . . . . . . G6 . . . . . . . G7 ESSDERC ’06, Montreux X1 X2 Partition A Partition B Event Driven: Partition (Matrix) B is evaluated only if event at net A happens Matrix A . G1 . . G2 . . . G3 ROBUSPIC Workshop Matrix B . . G4 . . . G5 . . . G6 . . . . G7 A. Baguenier – Slide 8 Multi Rate Simulation Low frequency partitions fin PFD High frequency partitions CP VCO f/N •• •• •• Highfrequency frequencypartitions partitionsrequire requiresmall smalltime timesteps steps High Lowfrequency frequencypartitions partitionsallow allowbigger biggertime timesteps steps Low MultiRate Rateenables enablesdifferent differenttime timestep/speed step/speedfor forpartitions partitions Multi ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 9 Representative Transistor Models DigitalMOSFET MOSFETModel Model Digital •• •• 20Xfaster fasterthan thanBSIM3 BSIM3 20X Accuratefor fordigital digitalbehavior behavior Accurate (speed,power) power) (speed, Optimizedfor forspeed speed •• Optimized AnalogTable TableModel Model Analog •• •• •• •• 5Xfaster fasterthan thanBSIM3 BSIM3 5X Accuratefor fordigital/analog/RF digital/analog/RF Accurate circuits circuits Optimizedfor foraccuracy/speed accuracy/speed Optimized Nonlinearcharge chargepreserving preserving Nonlinear model model ESSDERC ’06, Montreux We do the same for other MOS devices, for examples: MOS9, MOS11, PSP, etc…. ROBUSPIC Workshop A. Baguenier – Slide 10 Hierarchical Isomorphic Simulation X1 Split due to input change X2 X3 Adaptive Partitioning Subcircuitsseeing seeingthe thesame sameinput inputare aresimulated simulated •• Subcircuits byone onerepresentative representative(1 (1partition partitionfor for1000 1000 by inverters) inverters) Hierarchycan canbe bekept keptas aslong longas asthere thereare are •• Hierarchy circuitssharing sharingsame samerepresentative representative circuits instancesof ofsame samesubcircuit subcircuitsee seedifferent different •• IfIfinstances inputsignals, signals,new newrepresentative representativeisiscreated created input (splitpartition) partition) (split ESSDERC ’06, Montreux X2 X3 0V 0V XINV<999:0> X1 ROBUSPIC Workshop XINV<345:0> 3V XINV<999:346> A. Baguenier – Slide 11 MOSFET Modelling Spice Model D G I,Q S • Analytical model (Spice equations) used for IV and QV B V Automatic simplification Example for pure digital bloc D B ESSDERC ’06, Montreux ROBUSPIC Workshop D G B S S B A. Baguenier – Slide 12 Customized Partitioning Digital Appl. Analog/High Precision Appl. Aggressive Partitioning ADC DAC PLL VCO Conservative Partitioning Bitline Memory Appl. Wordline Aggressive Partitioning ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 13 IR Drop and Ground Bounce – Impacts timing and leads to failed silicon. – IR Drop (ground bounce) increases clock skew. • Hold time violations – IR Drop (ground bounce) increases signal skew. • Setup time violations – EM is based on average current density. – Grids have become so complex that current flow is not necessarily intuitive. VDD = 1.20V VDD = 1.1V CLK VDD = 1.17V ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 14 Virtuoso UltraSim Power -Net Solver Example Power-Net • Mixed-Signal Design – 3 supply voltages – 80% parasitic on power-net Element Count MOS DIODE R C Total 10K 256 249K 3.5k 268K Simulation time 5 ns transient Platform 2.6 GHz Linux ESSDERC ’06, Montreux UltraSim UltraSim Power-Net Solver 2 hrs 36.6 min ROBUSPIC Workshop A. Baguenier – Slide 15 Electromigration (EM) Transistor netlist n4 n3 Transistor current calculation n2 Vectors n1 n8 n5 n6 n7 Power-Grid Database Electromigration analysis EM Plotting – Use VoltageStorm to analyze EM on power grids. ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 16 Reliability A Growing Concern • HCI and NBTI Severely Impact Design – Wasted performance with excessive guard bands – Yield reduction due to immediate failures • Must be Addressed With Design Solutions • What is Reliability Simulation – Simulating the design to check the probability that a system will perform its functions over the required lifetime in the defined operating conditions without failure ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 17 Hot Carrier Effects • Emax at drain corner causes hot carrier generation • Hot carriers cause Isub, Igate and oxide damages S D G For NMOS & PMOS Ig n+ n+ n+ Isub Isub P-well ESSDERC ’06, Montreux Oxide Damage Impact Ionization ROBUSPIC Workshop A. Baguenier – Slide 18 NBTI a 0.13 Micron CMOS Crisis • Negative Bias Temperature Instability (NBTI) – – – – For PMOS at high temperature Driven by high vertical field Just appearing in tox < 5nm Even for zero Vds & long channel devices S G D p+ p+ • Causes immediate failures – Also degrades performance & yield of PMOS devices N-sub Isu b Vdd Mainly PMOS ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 19 Reliability Simulation Flow With UltraSim • Built-In support for HCI and NBTI • Provides reliability information – – – – Device aging information Lifetime degradation Fresh simulation results Aged model simulation results • The age reliability of the RF performance of CMOS radio chips are validated prior to high volume manufacturing see ref[1] ESSDERC ’06, Montreux RelPro+ BSIMPro+ Reliability Spec Reliability Parameters Models Netlist UltraSim • Device Age Information • Lifetime degradation • Fresh vs. Aged waveform Comparison ROBUSPIC Workshop Fresh vs. Aged A. Baguenier – Slide 20 AMS Top Level Tool Flow Simulation Strategy Dev System Level IP From System/IC Flow VSDE (ADE) To AMS Block Creation Flow To RFIC Flow HDL Sim AMS Designer RC Parasitics From Chip Integration Flow Mixed Level Simulation AMS Designer Spectre Ultrasim NC Hierarchy Editor Uses UltraSim in AMS simulation Transistor Level Block Descriptions From AMS Block Creation Flow ESSDERC ’06, Montreux Calibrated HDL Models From AMS Block Creation Flow ROBUSPIC Workshop RFIC Content From RFIC Flow A. Baguenier – Slide 21 Facing Reality • Verilog-AMS or VHDL-AMS? – Different languages have different strengths – We are not starting from scratch – Different people like different languages The Ultimate Solution? • ADA was the ultimate programming language • VHDL was the ultimate HW description language • SystemC was supposed to replace all other languages • PSL was supposed to be the ultimate assertion language ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 22 The Wrong Reasons It’s Fun! • Emotional • Religious • Controversial Verilog-AMS or VHDL-AMS ? Languages Are NOT The Main Issue • They are a means to an end • They enable methodologies • They interface to the engines ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 23 What Is The Main Issue? Getting first silicon (and software) out the door, shipping to customers, clean and on-time ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 24 VHDL -AMS VHDL-AMS • To insure the model compatibility some companies decide to not use SPICE primitive but convert each SPICE primitive example diode, bjt, bsim3 model to a VHDL-AMS code • Is it a good direction to be simulator independent? • In VHDL-AMS some key statements do not provide the error limit specification – wait on aRealQuantity'above(SomeRealSignal); • Do you start your model coding using the 1076.1.1 IEEE Standard VHDL Analog and Mixed-Signal Extensions—Packages for Multiple Energy Domain Support final version ?.... ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 25 VHDL -AMS & Verilog -AMS VHDL-AMS Verilog-AMS • Third generation simulators may have limitations for VHDL-AMS or Verilog-AMS model: – The automatic simplification may not apply on a BSIM3 coded in VHDL-AMS or Verilog-AMS model – The VHDL-AMS model implementation is not optimized for speed, like C++ internal model codes. – The partitioning… yes or may be …. • For new devices model development Verilog-AMS language seem the best languages, because of ADMS. – Be aware that ADMS support a limited subset of Verilog-AMS. – Study it before staring the model coding. • Conclusion: – You should not spend time to recreate the same SPICE model in Verilog-AMS & VHDL-AMS – I suggest to keep the SPICE primitive for fast simulation ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 26 Classical errors 1. Modeling error 1. Capacitor modeling, Model discontinuity, function differential (avoid nim, max, abs….) 2. 3. Check the Weff & Leff of your devices. Look to the voltage of the grounds. Floating grounds may slowdown your simulation. (no connection to the electrical solver ground) 4. An oscillator in time domain simulation starts because of the integration error created by your solver! – Different integration methods, Different simulator can provide different starting behavior. 5. Check the elements cut by the simulator. 6. Check the nodes cut by the simulator. 7. If you have a power up sequence, you should start the digital stimuli after the supply is established. 8. If you are using parasitic back annotation , turn on the RC reduction. 9. Do not reduce the parasitic from the power sensing function. 10. When you save the currents, the system of equations is different in the electrical solver. 11. Reliability check which Vgs & Vds bias are the worst condition for HCI & NBTI issues. 12. Read the simulator log files. ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 27 Unused Solver features capability • Use envelope analysis to speed up the SMPS simulation. • Use periodic steady state, for SMPS working at constant frequency. • Use the SpectreRF PXF for switching test case and stability testing for the feedback system, open loop analysis. ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 28 Simulation Results Analysis • • • With the new generation of software like AMS Designer simulator which include a FastSpice solver. Designers are able to speed up simulation: example a Complex Power Management circuit simulated in a SPICE like simulator during 3 months versus a 2 days simulation time in Ultrasim. For Smart power IC the current vectors are important for understanding the circuit behavior in addition to the voltages. • • More and more designers run simulation and save all currents. The result databases exceeding 30 giga bytes, now , become usual. • How to display such database results? • One solution for displaying such huge data base above 30Giga is to use CPU very good in IO performance. ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 29 Interesting CPU • The Sun Sun Fire V40z Server using Dual Core AMD Opteron 4 Dual Core AMD Opteron(tm) Processor 880 2390.780MHz, 32Giga of RAM You open a 30Giga simulation result database for 6 levels of hierarchy (V and I) and plot 20 electrical signals in less than 15 minutes.. The CPU IO performance is a key… ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 30 Conclusion • Power IC designs Flows are available. • Some improvements are still required in reliability modeling, PDK creation, and usage. • Designer education, on behavioral modeling, and tools usage remain. • Simulation technology continuous improvements are required for speed, capacity and huge data base display performance. • E language seems to be the future for System on chip verification. ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 31 Acknowledgements Many thanks to… • Christian Maier, (Robert Bosch) for the Bosch contributions to Robuspic WP4 • Renaud Gillon (AMI Semiconductor ) for the AMIS contributions to Robuspic WP4 ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 32 References • [1]: Mark Ruberto, RFIC Design Group, Wireless Products Division, Intel Israel Design Center, Haifa, Israel, 2005 IEEE Radio Frequency Integrated Circuits Symposium, Consideration of Age Degradation in the RF Performance of CMOS Radio Chips for High Volume Manufacturing • [2]: The Future Verification Environment: The Multi-Language Approach, Cadence Design Systems presentation. • [3]: ADMS - Automatic Device Model Synthesizer, Laurent Lemaitre’, SPS – Motorola, IEEE 2002 CUSTOM INTEGRATED CIRCUITS CONFERENCE • [4] Virtuoso® UltraSim Simulator User, Cadence Design Systems. ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 33 Thank you ESSDERC ’06, Montreux ROBUSPIC Workshop A. Baguenier – Slide 34