Cognitive Information Processing Technology Brief to High Performance Embedded Computing Workshop 2002 Mr. Zachary J. Lemnios Information Processing Technology Office Deputy Director 24 September 2002 Acknowledgements • • • • • • • • • • Ron Brachman Bill Dally Bob Graybill David Honey Mark Horowitz Steve Keckler Dave Koester Bob Leheny Christie Marrian Dan Radack DARPA/IPTO Stanford University DARPA/IPTO DARPA/ATO Stanford University University of Texas Mitre DARPA/MTO DARPA/MTO IDA Page 2 Agenda • Introduction – DoD System Challenges – DARPA Offices and Programs • Technology Trends – Device Performance – High Performance Computing • Cognitive Processing Technology • Summary Page 3 Asymmetric Advantage Enabled by Information Superiority See Further with Greater Clarity Space Based RADAR E-3 AWACS Airborne Early Warning FLTSAT Secure Comm GMTI, SAR, STAP, HSI RC-135V Rivet Joint Space-Time Adaptive Processing SIGINT E-2C Hawkeye Tier II+ UAV Standard Missile Cooperative engagement Aegis Cruiser Towed Sonar Array Adaptive Matched-Filter processing TEL Wideband high linearity target discrimination multi-angle target discrimination Chemical biological threat detection Small Unit Operations Covert sensing Power constrained operations Secure LPI A/J communications Page 4 Information Networks Will Revolutionize Platforms TODAY - Federated Architecture (Baselines 1-5) Point-to-Point Mainframes (UYK-7/43) Limited Growth Capability Vulnerable to Damage FUTURE - Distributed Processing (Future B/L) Highly Distributed Network Redundancy Plus Reconfigurability Effectively Invulnerable to Battle Damage The Upside: Agile, Adaptable, Survivable Systems The Downside: Complexity in h/w, s/w, integration From: www.hypocrites.com/pictures Page 5 A New Class of Autonomous Systems Mission Complexity Autonomous Systems UCAR (Artist’s concept) OAV (Artist’s concept) UCAV UCAV-N Spinner WV (w/ 390kg fuel) (w/ 20kg for nav) (w/ 345kg battery) OAV (w/60kW generator) (Open Terrain) HMMWV (w/ 390kg fuel) (w/ 20kg for nav) (w/ 345kg battery) (w/ 60kW generator) TMR Weight w/fuel,sensors (No payload) 4,893kg (10,764lb) UGCV 3,736kg (8,220 lb) 1,500kg (3,300lb) (Artist’s concept) 945kg (2,080 lb) 6,393kg (14,064lb) 4,672kg (10,300 lb) Payload fraction 24% 20% Vertical obstacle 1.2m (47") 0.56m (22") Trench 2.1m (80") 0.8m (31") 0.42m (16") 0.45m (18") 70% (35°) 70% (35°) 108 kPa (15 psi) ~220 kPa (32 psi) Global Hawk Cooperative Systems Legacy Systems A160 Payload capacity (after imposed loads) GVW with full payload Power (peak) 192kW (257 hp) 139kW (186 hp) • Human-like sensor depth Power/Weight ratio (peak) 43 hp/ton 39.8 hp/ton • Persistent engagement CG height / Track ~0.40 ~0.59 Vertical center-of-gravity 0.8m (31") constraints 1.1m (43") • Extreme power and volume Width 2.5m (100") 2.3m (90") • Effects-based tasking and performance Tire diameter 1.2m (47") 0.95m (37") Environmental Complexity Ground clearance Side slope Min ground pressure Fording Max road speed 1.2m (47") ~62kph (38mph) 0.9m (35") Page 6 125kph (80mph) DARPA Organization Director, Tony Tether Information Exploitation Richard Wishner Steven Welby / Robert Tenney Sensors Exploitation Systems Command & Control Information Awareness John Poindexter Robert Popp Asymmetric Threat Prediction Behavior Modeling Tactical Technology Allen Adler Art Morrish Air/Space/Land Platforms Unmanned Systems Space Operations Laser Systems Future Combat Systems Planning / Logistics Special Projects Amy Alving Joe Guerci Chem/Bio Def Systems Counter Underground Facilities Space Sensors/Structures Navigation/Sensors/ Signal Processing Advanced Technology Tom Meyer Dave Honey Assured C3ISR Maritime Early Entry/Special Forces IA&S Programs Defense Sciences Michael Goldblatt Steven Wax Bio Warfare Defense Technologies Biology Materials & Devices Mathematics Information Processing Technology Ron Brachman Zach Lemnios Cognitive Systems Computational Perception Representation & Reasoning Learning Natural Communication MicrosystemsTechnology Robert Leheny Electronics Optoelectronics MEMS Combined Microsystems Focus of this brief Page 7 DARPA/MTO Platform Scale Information Systems Process Sense Actuate Memory RF w MMW IR UV Controller Sensor(s) Switching Analog Digital Visible Bio I/O HumanInterface Interface Human MachineInterface Interface Machine WeaponInterface Interface Weapon NetworkInterface Interface Network Output Output A/D A/D Processor(s) • Highly capable sensors- sensors that are self adapting • Enhanced extraction of signals from background, noise, and jamming • Covert “data” into actionable “knowledge” in near real time • Provide technology for assured communication links Page 8 The Challenge of Complexity 1.E+11 1.E+10 Ops/sec/$ doubles every Few months 1.E+09 1.E+06 Tubes/ Transistor 1.E+03 1.E+00 1.E-03 Mechanical/ Relays doubles every 7.5 years doubles every 1.0 years nanometer CMOS doubles every 2.3 years Combination of Hans Moravac + Larry Roberts + Gordon Bell 1.E-06 1880 1900 1920 1940 1960 1980 2000 2010 2020 2030 While computational performance is increasing, productivity and effectiveness are not keeping up: – Users must adapt to system interfaces, rather than vice versa – Systems have become more rigid and more fragile – Systems have become increasingly vulnerable to attack – The cost of building and maintaining systems is growing out of control Page 9 DARPA/IPTO Cognitive Systems • DARPA IPTO will create a new generation of cognitive computational and information systems with capability to: – reason, using substantial amounts of appropriately represented knowledge – learn from their experience so that they perform better over time – explain themselves and be told what to do – be aware of their own capabilities and reflect on their own behavior – respond robustly to surprise Systems that know what they’re doing Page 10 Why Now? • Human-level scaling of HW technology is on the horizon • Foundations established for human neural systems • Cognitive technology (from AI) is being applied to initial problems Page 11 The Result Will Enable a Revolution in Capability for DoD More Aggressive Threats Adaptive and Intelligent Data-Fused Sensors Threats are more dynamic and in deeper hide (collapsing time lines) System performance is outpaced by changing threat environments Cooperative battle management requires robust information backbone Sensor Data Flow Overwhelming Human Analyst Cognitive Information Exploitation Sensor bandwidth is increasing faster than processor capability Target classification has become a multi sensor The next revolution in sensing: Autonomous Adaptation The next revolution in computing: Cognitive Processing Page 12 Agenda • Introduction – DoD System Challenges – DARPA Offices and Programs • Technology Trends – Device Performance – High Performance Computing • Cognitive Processing Technology • Summary Page 13 Beyond CMOS: The Road Beyond Bulk Silicon Field Effect Transistors 10.00 Normalized Device Speed ITRS Roadmap Projections Quantum dots/Cell. Automata Self-directed assembly metallic particle SiO2 Bulk-Si Best Case Predictions SET IEDM Benchmark Technologies 1985 1990 1995 2000 2005 2010 D 2015 G 2020 G Molecular devices S S Courtesy Dimitri Antoniadis, MIT; Rob Rutenbar, CMU silicide Gate Oxide Si Gate B Nonplanar Switches 1/size Nanotubes nFET Gate { 0.10 NDR, RTD { 1.00 pFET Silicon Schottky barrier isolation Schottky source/drain FET Si (tensile) Bulk CMOS Double-Gate CMOS Si0.8Ge0.2 High k gate oxide Metal gate After Chart by P. Wong, IBM Si1-xGex Si SiGe/Ge FETs and Structures (strained layers) Page 14 Memory Wall is Growing 100 Memory access times (cycles) are increasing based on SIA clock frequency roadmap 180nm 130nm 100nm 70nm 50nm 10 1 1 10 100 1000 10000 Cache Capacity (KB) Source: Keckler – Univ. of Texas Page 15 SIA Roadmap Impact on Computer Architectures 0.28 0.26 • 400 mm2 Die 700 MHz 0.24 0.22 Process (microns) 0.2 Single Clock Area 1.25 GHz 0.18 0.16 2.1 GHz 0.14 0.12 0.1 6 GHz 0.08 10 GHz 0.06 13.5 GHz 0.04 0.02 0 1996 1998 2000 2002 2004 2006 Year 2008 2010 2012 2014 New architectures are required to accommodate smaller clock regions Page 16 Novel Architectures are Required to Extend Performance Productivity 1e+7 Perf (ps/Inst) Delay/CPUs 1e+6 1e+5 1e+4 1e+3 1e+2 30:1 1e+1 1e+0 1e-1 1e-2 30,000:1 Opportunity for Cognitive Architectures 1e-3 1e-4 1980 1985 1990 1995 Source: ISAT Summer 2001 Study- Last Classical Computer; Prof. Bill Dally (Stanford U) Study Lead 2000 2005 2010 2015 2020 Page 17 Embedded Computing Performance Regions 10000 Computation Density (MOPS/cm3) Custom VLSI 1000 Polymorphous Architectures 100 Reconfigurable Programmable 10 1 0.1 0.01 0.1 1 10 100 1000 Computational Efficiency (GOPS/Watt) Page 18 Polymorphous Computing Architectures Program MultiMission Multiple Sensors (A,B,C...X) plug & play A B C…. X In-Mission Re-target -ability Platform Tracking transit Multi-sensor processing Goal: Computing systems (chips, networks, software) that will morph to changing missions, sensor configurations, and operational constraints during a mission or over the life of the platform. P e r f o r m a n c e PCA Morph Space Mission Selectable Virtual Machines Architecture Space Specialized DSP Class Class PPC Class Server Class Mission Aware Embedded Computing Page 19 Agenda • Introduction – DoD System Challenges – DARPA Offices and Programs • Technology Trends – Device Performance – High Performance Computing • Cognitive Processing Technology • Summary Page 20 DARPA/IPTO Cognitive Systems • DARPA IPTO will create a new generation of cognitive computational and information systems with capability to: – reason, using substantial amounts of appropriately represented knowledge – learn from their experience so that they perform better over time – explain themselves and be told what to do – be aware of their own capabilities and reflect on their own behavior – respond robustly to surprise Systems that know what they’re doing Page 21 Cognitive Systems Thrusts Dynamic Coordinated Teams Systems Systems That Know What They’re Doing Core Cognition Cognitive Architecture Perception Representation & Reasoning Learning Communication & Interaction Robust Software and Hardware Foundational Science and Mathematics Bio-inspired Computing, new approaches to Trust Management Foundation Page 22 Cognitive Agent Anatomy of a Cognitive Agent Perception Reflective Processes LTM STM Deliberative Processes Other reasoning Communication (language, gesture, image) (knowledge base) Concepts Sentences Prediction, planning Action Reactive Processes Sensors Effectors External Environment Page 23 Initial Challenge Context Persistent, personal partner/associate systems • Learn from experience • Learn what you like and how you operate by observation by direct instruction or guidance, in a natural way • Imagine possible futures, anticipate problems and needs • Omnipresent / always available Examples • • • • Commander’s (C2) assistant (Intelligence) Analyst’s associate Personal executive assistant/secretary Disaster response captain’s “RAP” (robot/agent/person) team Page 24 Summary DoD is facing immense challenges • New and dynamic threats in much deeper hide • Collapsing timeliness, rapidly changing threat environments New classes of autonomous systems require • Platform Scale Integration (DARPA/MTO) • Cognitive Capabilities (DARPA/IPTO) Systems that Know what they are doing • can reason • can learn from their experience • can explain themselves • can be aware of their own capabilities • can respond robustly to surprise Page 25 Challenge •Send us your best ideas: IPTO BAA 02-21, http://www.darpa.mil/ipto/Solicitations/PIP_02-21.html MTO BAA Solicitation http://www.darpa.mil/mto/solicitations/index.html •Take a tour as a DARPA Program Manager rleheny@darpa.mil rbrachman@darpa.mil zlemnios@darpa.mil (703) 696-2268 (703) 696-2264 (703) 696-2234 Page 26 Backup Page 27 Memory Issues Lead to Inefficient Performance Year of Introduction (Cray) 1988 1994 2000 Cray C90 Y-MP T90 J90 Intel Alpha SGI MIPS Sun Cray EL-90 10.00% SV1 Microprocessors Measured %peak (STREAMS ADD) 100.00% 1.00% 10 100 1000 10000 Clock Speed MHz STREAMS ADD: Computes A + B for long vectors A and B (historical data available) Page 28 Today’s Systems - Collection of Rigid Embedded Subsystems Static Mission Scripts System-of-Systems Micro-Systems SubSubSystem System Sense ………. SubSystem Micro-Systems SubSubSystem System SubSystem Sense Process Actuate Process Actuate Single Instance, Point-Design Implementations Mode Driven Configuration Predefined Functional Capability DoD Systems must Move from Integration Driven Architectures to Capability Driven Architectures Page 29 High Performance Embedded Cognitive Systems System-of-Systems Cognitive Information Extraction Cognitive micro-Systems SubSubSystem System Sense ………. SubSystem Process Actuate Cognitive micro-Systems SubSubSystem System Sense Local Hard Real-Time Cognitive Agents SubSystem Process Actuate Goal(s) Driven Missions Cognitive Guided Cognitive “Enabled” Agile Subsystem (e.g. PCA) Systems with Human Like Capability Page 30 Intelligent Systems for Mission Agility Cognitive Processing & Adaptation Resource Aware Resource Aware Resource Aware u Sense Process Actuate Systems with Human Like Capability Page 31 DARPA/IPTO Cognitive Systems • While computational performance is increasing, productivity and effectiveness are not keeping up – – – – Users must adapt to system interfaces, rather than vice versa Systems have become more rigid and more fragile Systems have become increasingly vulnerable to attack The cost of building and maintaining systems is growing out of control Systems That Know What They’re Doing Perception Representation And Reasoning Learning Communications And Interactions Robust Software and Hardware Foundational Science and Mathematics (incl. Bio-inspired Computing, new approaches to Trust Management,…) Page 32 Information Technology • Information Exploitation Office – Detection, Precision ID, Tracking, and Destruction of Elusive Surface Targets • Information Awareness Office – Detect & Defeat Terrorist Networks • Information Processing Technology Office – Cognitive Systems • Advanced Technology Office – Robust, Secure Self-Forming Tactical Networks Page 33 Feature Size Trends Intel8080 Feature size (nanometers) 1000nm Intel386 Intel486 Pentium PentiumPro 100nm Pentium IV Synchronous (clocked) DARPA 10nm 1nm 1970 ? 1980 1990 2000 CMOS 2010 2020 2030 “Beyond CMOS” 2040 2050 Page 34 High Productivity Computing Systems Program Goal: Provide a new generation of economically viable high productivity computing systems for the national security and industrial user community (2007 – 2010) Approach: Implement productive high-end systems with high effective bandwidth, low latency, balanced system architecture, robustness, application responsive tailorability, performance measurement and prediction Applications: Intelligence/surveillance, reconnaissance, cryptanalysis, weapons analysis, airborne contaminant modeling and biotechnology Fill the Critical Technology and Capability Gap Today (late 80’s HPC technology)…..to…..Future (Quantum/Bio Computing) Page 35