Virtual Algorithm Design for RF Receivers Sanjay A. Khan k.inc k.inc proprietary, all rights reserved 1 Wireless Digital Communication Upconversion Downconversion • Noise removal at receiver provides opportunities for more aggressive utilization of bandwidth • Sources of “noise”: - mixing in of adjacent carriers due to circuit non-linearities - thermal and shot noise in receiver - phase noise introduced by frequency synthesizer in receiver - substrate noise from digital circuitry in receiver • Only thermal and shot noise in receiver are truly stochastic with no autocorrelation • All other sources of noise are “deterministic” and can be theoretically removed at the receiver • Shot and thermal noise are the only truly “unremovable” impairments. Orders of magnitude improvements in efficiency that could be made possible with infinite digital processing power at receiver k.inc proprietary, all rights reserved 2 Wireless Communication Landscape k.inc proprietary, all rights reserved 3 Next Generation Wireless Communications Standards Next Generation Wireless LANs: IEEE 802.11n • • Raw throughput between 135Mb/s and 540Mb/s Pre-standard products emerging now (2006 standards) Wireless Personal Area Network: • • Wireless USB (WUSB): 802.15.3a Up to 480Mb/s throughput with a range of 10 meters Look for WUSB to appear in laptop chipsets in 2006 Broadband Wireless Access / Metropolitan Area Networks: • • • • IEEE 802.16 WiMAX Throughput up to 70Mb/s with a range up to 20 miles Licensed spectrum (2.5-2.6 GHz) or Unlicensed (2.4 & 5.8 GHz) 802.16a products emerging now 802.16e provides support for mobility and handoffs (2006) IEEE 802.20 MobileFi • • • Throughput up to 12Mb/s with a range < 10 miles Supports mobile users up to 250 Km/h in licensed spectrum bands Some pre-standard technologies being fielded (e.g., Flarion.Nextel) k.inc proprietary, all rights reserved 4 Effect of Impairments on QAM Modulated Signals . . . . . . . . Q . . . . I . . QAM 16 Constellation RF signal transmitted: I cos(RF t ) Q sin( RF t ) . . Shannon’s Channel Capacity Formula C 2 B log 2 (1 S ) N For a given baud rate (channel bandwidth), higher SNR allows more aggressive constellations, and consequently higher bits per second k.inc proprietary, all rights reserved 5 Opportunities to Algorithmically Improve Spectral Efficiency Infinitely powerful . . . . . . . . . . . . . . . . Baseband DSP Need for high fidelity RF circuit simulator that can capture all RF impairments at circuit Level accuracy . . . . . . . . . . . . . . . . while co-simulating with Baseband simulator • Compression due to non-linear gain can be statically compensated • Adjacent channel mixing impairments can be equalized out if receiver is powerful enough to detect symbols from all adjacent channels and then construct an equalizer • Symbol rotation due to PLL is autocorrelated, so is theoretically cancelable • Substrate Noise effects can be detected and nulled out with clever design • Beam steering technologies allow higher SNR • Uncorrelated noise introduced by the LNA and downconversion mixer is unremovable. This is the floor noise level is determined by temperature and the Boltzmann constant and cannot be equalized out. k.inc proprietary, all rights reserved 6 Product Optimization Via Virtual ProtoTyping Many, many opportunities for clever, optimal design of RF transceivers – perhaps two decades of innovation to come Product roll-out for future generations must follow an evolution trajectory that is consistent with what consumers will incrementally pay for incremental performance k.inc proprietary, all rights reserved 7 Value Proposition • Current wireless technologies use the precious resource of bandwidth in an abysmally inefficient way (WCDMA/UMTS, WiFi, FlashOFDM) RFIC Vendors need to differentiate themselves by trying to achieve: • The least amount of compression, rotation and symbol spread in the received constellation (for higher bit rates) • at the lowest cost and power Wireless Food Chain • • • • • • Worldwide Wireless Service Provider Market Worldwide RF Communication Equipment Market Worldwide RF Communication IC Market Worldwide RF Transceiver IC Market CAD for Xcvr Alg. Exploration (BB Alg. Sim, Ckt Extraction, RF Ckt Sim) Accurate RFIC Circuit Simulation (key enabler for virtual proto-typing) $1 Trillion $300 Billion $50 Billion $20 Billion $1 Billion $100 Million Real proto-typing is extremely expensive ($millions per pilot run), turn around is painfully slow (many weeks), and it does not provide algorithm designers an opportunity to “play” with quick feedback $100 million opportunity to provide wireless xcvr algorithm designers with a cheap(er) virtual sandbox that they trust. Incorrect algorithm optimization due to poor simulation fidelity is a big pain for system architects PAIN KILLER: A fast, highly accurate virtual proto-typing tool that can efficiently handle all components of an RFIC (Power Amp, Upconverter, LNA, Downconverter, VCO, Phase Detector, Frequency Dividers, Filters), with all impairments (ckt element noise sources, gain compression, adjacent channel mixing, PLL phase noise, substrate noise) simulated at circuit level accuracy k.inc proprietary, all rights reserved 8 Sources of RF Signal Impairments PA Low Noise Amplifier • High gain, high sensitivity • Non-linearities cause blockers to mix into channel of interest => spread in rcvd constellation • Non-linear gain causes compression of rcvd constellation • Component Noise Power (Thermal / Shot) is comparable to power of rvcd signal - causes spread in rcvd constellation Air Interface LNA (attenuation + blockers) Power Amplifier • Non-linear (class E) • Non-linearities cause symbol spread due to mixing in of adjacent channels (APCR) • Non-linear gain causes compression of rcvd constellation RF UpConv BB (I/Q) Downconversion Mixer • LO Phase noise causes angular spread in rcvd constellation • Component noise causes spread in rcvd constellation (cyclostationary) DnConv Rcvd Constellation Automatic Gain Control • Non-linearities causes symbol spread due to mixing in of adjacent channels • Gain Compression causes compression of rcvd consellation LO LO PLL • VCO phase noise causes rotation of rcvd constellation • PLL dynamics cause angular spread in rcvd constellation AGC Loop Filter Counter VCO Frequency Synthesizer k.inc proprietary, all rights reserved 9 Competitive Landscape • • • • • • No vendor offers a circuit simulation tool that does end-to-end symbol in / symbol out simulation of the entire RF system at the circuit level, accounting for all sources of noise. For standalone transmitter (Power Amplifier), several vendors offer time domain symbol in / symbol out simulators using Envelope Transient Harmonic Balance. Agilent, Xpedion and AWR offer the most widely used solutions. K.inc’s solution will be faster than any of the above. Berkeley Design Automation offers circuit level tools to simulate phase noise in standalone VCOs and standalone open loop PLLs. No vendor offers a circuit level tool that simulates the entire frequency synthesizer. Agilent’s Envelope Transient algorithm is inappropriate for PLL circuits. No vendor offers a time domain solution for the low noise amplifier and downconversion mixer which accounts for phase noise in the LO (freq. synthesizer) and component noise in the LNA and mixer. Berkeley Design Automation’s technical team have No efficient circuit level simulation tool is available for the AGC. In lieu of the non-availability of circuit level time domain simulation tools, many vendors offer tools to create time domain behavioral models of the above circuit components after pre-characterizing them with frequency domain circuit simulation (Cadence JK Models, Agilent and AWR with AM/AM and AM/PM models). Xpedion offers a Neural Network based tool to create behavioral models. All the above behavioral models have highly questionable and suffer from mathematically unbounded inaccuracy. k.inc proprietary, all rights reserved 10 k.inc Technologies for Robust RFIC Circuit Simulation • • • • • • • • • Device models (like BSIM4) re-derived using Automatic Differentiation for robustness Table models with speed / accuracy trade-offs appropriate for different portions of the system. Third order derivative accuracy for RF mixers and amplifiers, first order accuracy (at considerably higher speed) for “large signal” circuits. To be patented. Homotopy algorithms (based on LOCA) for robust DC convergence and Autonomous Periodic Steady State computation. Multi-rate Partial Differentail Equation (MPDE) Circuit Solver with Shur Complement matrix algorithms and Volterra Series techniques to allow lightning fast simulation of RF transmitters. To be patented. LTI (Linear Time Invariant) simulation techniques with Volterra Series corrections for high speed simulation of LNAs and downconversion mixers. Novel algorithm to rigorously handle phase noise introduced by LO; Ornsten-Uhlenbeck algorithm for time domain noise simulation. To be patented. Fokker Plank solver for characterizing phase noise in VCOs. Fast Timing circuit simulation algorithms for handling other blocks of PLL. To be patented. Adaptive Volterra Series algorithm for simulating AGCs. To be patented. All numerical computation tuned to exploit hyper threading and other features of modern processors; custom for device models, based on widely used open source packages for matrix computations. System simulation set up to exploit multi-threading and multiple workstations on a LAN Extensive use of robust software packages (PVM, MPI, PETSC, UMFPACK, ADOL-C, LOCA, QT). k.inc proprietary, all rights reserved 11 Product Positioning c S ircuit imulator for h a n n igh frequency nalog R oise-limited non-linear circuits and o ptimization sandbox for n on-linearities and noise k.inc proprietary, all rights reserved 12 Bottom Up Market Analysis • Wireless Standards needing circuit-level symbol in / symbol out simulation capability: - WiFi/WWiSE (802.11x), WCDMA/UMTS, 4G, Flash OFDM, WiMax, 802.16x, Bluetooth, UWB • Potential customers engaged in design of RFICs for above standards: - Agere, Airgo, Airoha, Analog Devices, Atheros, Bermai, Broadcom, Cambridge Silicon Radio, Conexant, Envara, Freescale, GCT, IceFyre, Infineon, Intel, Intersil, Marvell, Mobilian, Parthus, Philips, PlexTek, Qualcomm, Railink, RFMD, Signia Tech, Silicon Wave, ST Micro, Synad, Systemonic, TI, Wireless Interface Tech, Wiscom, Zeevo, Zyray, … • RFIC Market projected to grow to $37 Billion by 2006, with 663 projected design starts per + year over all the above standards . • Each design start will need one team license for cShannon. • cShannon will be licensed for $100,000 per design team per design start • Addressable market is approx. $70 Million per year! + k.inc proprietary, all rights reserved Cahners InStat, 2003 13 Exit Strategy NYSE CDN • • • • • • $1.13B revenue Xpedion - not yet public • Significant player in analog and mixed signal NSDQ SNPS Investors: verification markets (Virtuoso) • $1.18B revenue - Menlo Ventures Berkeley Design Automation • Declining presence in RFIC extraction • Significant player in and analog and mixed signal Applied Wave Research - Telesoft Partners • Investors: simulation hence technology partnership with simulation markets • Investors: - Telos Venture Partners - Bessemer Venture Partners Agilent • Ckt. Sim. Technology via acquisitions - CMEA Venture Partners - Redwood Venture Partners - Woodside • Homegrown Ckt. Sim. Technology:(Hspice) - MetaSoftware - Synopsys ? Fund Spectre - Anagram First Round: $12M First Round: $13M Spectre RF (outdated - Epictechnology) First Round: Second Round: ? $5M • Ckt. Sim. Technology via acquisitions • Extraction technology from Avanti acquisition • No revenue yet BLDA (ADS) • No presence in RFIC verification market ? revenue ?• revenue • Very little information available on what circuit Celestry (Ultrasim) • Significant investor in AWR • Envelope Transient HB + frequency domain RFIC freq. domain circuit level simulation + NSDQ MENT NYSE A NSDQ LAVA simulation and analysis technolgies are offered. • Significant investor in Xpedion VolterraTransient Series simulation (Microwave Office) Envelope HB for Power Amps (Golden • •$692M revenue • Test equipment vendor $127M revenue CTO is Amit Mehrotra whoindustry has done prior work • Extraction technology from partner, Gate) • •Minor player in line analog and mixed signal • Presence in RF- not EDA publicly due to Eesof acquisition Emerging full EDA vendor Aplac traded in US in Fokker Planck based simulation of phase NSDQ ANST NoOEA extraction technology simulation markets • Market leader in frequency domain RF circuit • noise No presence in simulation market (digital • Pure RFIC player in oscillators. Other TAB members • Behavioral Simulation engine (Visual System/ • $57M revenue Behavioral Simulation using Neural Network • Weak RFIC ckt simulation technology (ELDO) simulation (ADS) simulation (Fast RF IC analog / RF) • RFIC circuit (Jaideep Roychowdhury, U. third of Minn.) Simulator) • RFIC freq.level domain circuit level simulation Models (Model Compiler) - uses partyhave • •Strong player in extraction (Calibre) • Strong in RFIC extraction (IC-CAP) No significant presence in extraction market Module) worked in macro-modeling of RF components. (Nexxim) behavioral simulator player in RFIC verification • •Not significant • Cadence partner for RFIC design; strong Mayaextraction look to filltechnology out product line in these areas • RFIC extraction (Electromagnetic Module) No offered. • RFIC extraction (Q3D) + electromagnetic integration with Cadence(System RFIC design flow • Behavioral Simulation Module) solver (HFSS) • Behavioral Simulation (Ansoft Designer) RFIC Simulation + Analysis Vendors Full Line RFIC EDA Vendors (Extr. + Sim.) Full Line EDA Suppliers Technology Supplier k.inc proprietary, all rights reserved 14 Development Status and Test Plan • All coding completed - Matrix computations - MPDE, LTI, Fokker-Planck, Ornstein-Uhlenbeck, Homotopy for PSS & DC Conv., Fast Timing - Automatically differentiation of non-linear device models, table models, Volterra Series stamps - distributed computation and message passing • • 125,000 lines of new C++ code (not including any packages). Testing milestones: - Validation of automatically differentiated models - Validation of first and third order table models - Time domain transient simulation - DC convergence with LOCA homotopy - LOCA homotopy for autonomous periodic steady state - Volterra correction stamps - MPDE with Volterra series and Shur Complement - Fast timing transient algorithm - LTI algorithm for LNA + downconversion mixer - Ohrnstein-Uhlenbeck noise simulation algorithm - Fokker Plank based oscillator phase noise algorithm - PVM set-up for multi threading and distributed computing k.inc proprietary, all rights reserved 15 Angel Funding Needs • • • • 6 man-months to complete coding 12 man-months to prove in one end-to-end design from friendly customer Total initial seed - $250,000 Expectations for final valuation – 33% of a $70 million market – Net profitability at steady state with 50% margin: $10 million per year – “Market Cap” at steady state: $100 million, assuming PE of 10 – Angel offered 10% equity in company, for 40X return on investment if exit on full maturity – Angel’s ROI reduced for early exit via acquisition, but still 10X even if eventual buyer gets 4X discount to offset higher risk of early acquisition k.inc proprietary, all rights reserved 16