Beyond von Neumann Dr. Mark E. Dean, PhD Fisher Distinguished Professor UTK College of Eng. 2009 IBM Corporation
Businesses are dying of thirst in an ocean of data 90% of the world s data was created in the last two years 80% of the world s data today is unstructured 20% is the amount of available data traditional systems leverages 2 1 in 2 business leaders don t have access to data they need Source: GigaOM, Software Group, IBM Institute for Business Value" 83% of CIOs cited BI and analytics as part of their visionary plan 2.2X more likely that top performers use business analytics
Our Connected World will Drive the Creation of Big/Fast Data Number of Connected Devices 50 50 Billion 40 30 20 15 Billion 10 7 Billion 2010 2015 2020 Multiple Sources: Intel, Ericsson, Gartner, etc.
New Big/Fast Data Brings New Challenges & Opportunities, Requires New Analytics Exa Homeland Security 600,000 records/sec, 50B/day 1-2 ms/decision 320TB for Deep Analytics Peta Up to 10,000 Times larger Telco Promotions 100,000 records/sec, 6B/day 10 ms/decision 270TB for Deep Analytics Data Scale Tera Giga Data at Rest DeepQA 100s GB for Deep Analytics 3 sec/decision Mega Traditional Data Warehouse and Business Intelligence Kilo Data in Motion Up to 10,000 times faster yr mo wk day hr min sec ms µs Occasional Frequent Real-time Smart Traffic 250K GPS probes/sec 630K segments/sec 2 ms/decision, 4K vehicles Decision Frequency
Analytics toolkits will be expanded to support ingestion and interpretation of unstructured data, and enable adaptation and learning New Data Traditional New Methods Adaptive Analysis Continual Analysis Optimization under Uncertainty Optimization Predictive Modeling Simulation Forecasting Alerts Query/Drill Down Ad hoc Reporting Standard Reporting Entity Resolution Relationship, Feature Extraction Annotation and Tokenization Responding to context Responding to local change/feedback Quantifying or mitigating risk Decision complexity, solution speed Causality, probabilistic, confidence levels High fidelity, games, data farming Larger data sets, nonlinear regression Rules/triggers, context sensitive, complex events In memory data, fuzzy search, geo spatial Query by example, user defined reports Real time, visualizations, user interaction People, roles, locations, things Rules, semantic inferencing, matching Automated, crowd sourced Ø Learn In the context of the decision process Ø Decide and Act Ø Understand and Predict Ø Report Ø Collect and Ingest/Interpret Decide what to count; enable accurate counting Extended from: Competing on Analytics, Davenport and Harris, 2007 5
The fourth dimension of Big Data: Veracity handling data in doubt Volume Velocity Variety Veracity* Data at Rest Data in Motion Data in Many Forms Data in Doubt Terabytes to exabytes of existing data to process Streaming data, milliseconds to seconds to respond Structured, unstructured, text, multimedia Uncertainty due to data inconsistency & incompleteness, ambiguities, latency, deception, model approximations * Truthfulness, accuracy or precision, correctness 6 6
By 2015, 80% of all available data will be uncertain Global Data Volume in Exabytes 9000 8000 7000 6000 5000 4000 3000 2000 1000 0 100 90 80 70 60 50 40 30 20 10 Aggregate Uncertainty % Data quality solutions exist for enterprise data like customer, product, and address data, but this is only a fraction of the total enterprise data. Multiple sources: IDC,Cisco By 2015 the number of networked devices will be double the entire global population. All sensor data has uncertainty. The total number of social media accounts exceeds the entire global population. This data is highly uncertain in both its expression and content. Enterprise Data 2005 2010 2015 7 7
Advances in computational systems 5-10 years Cognitive Computing Compute+ Natural Language+ Analytics Program Deep Q&A Computers 1 Big Data Synapse devices L e a rn BIG/Fast 00 X 0, 0 0 0, 1,000 à Data + analytics (zettabytes + milli / microseconds net Smarter Pla ple) ings + Peo h T f o t e rn (Inte Exascale (Datacenter-in-a-box) Massive parallelism Flexible system optimization Processing in-memory 1000X Workload Optimized Systems Nano Systems (Systems-on-a-chip) 1B Transistors Nano Devices (Power7 chip) 1000X 1T Devices Photonics DNA Transistor PhaseChangeMemory Carbon Nanotubes Memristor
Architectural Enhancements Relative Performance (log) Scalar processing RISC, CISC, Vector, (Frequency) Superscalar Out-of-Order Superpipeline Mutli-level Caches (Freq. & memory) Workload optimized - Petaflop - Deep Q/A - Storage - SSD Multicore SMT Processor Networks (Chip Density) 1970 2000 2018
Log of Compute Power 1E+12 Integrated Circuit Nanotechnology $1000 Buys: Computations per second 1E+9 1E+6 1E+3 1E+0 1E-3 Mechanical Electro- Mechanical Vacuum Tube Discrete Transistor 1E-5 1900 1920 1940 1960 1980 2000 2020 Source: Kurzweil 1999 Moravec 1998
Device Structure Research Pipeline C Electronics Fully Depleted Devices HfO2 ETSOI Si NW Deposited Si FINFET Conventional Planar Device 22/20 nm 15/11 nm 8 nm & Beyond Si Nano-Wire
The Charge to Exascale: Future Technologies 1 PetaFlop 72 BG/P Racks Overall Performance = 1000X Overall Accessable Data = 10^6X Performance / watt = 135X Performance / $ = 1000X Footprint = <2% Referenced to1pf system CPU Phase Change Memory Software Silicon Photonics 3D CNT Graphene The Next Ten Years 10 PetaFlop 100 P7IH Racks 1 PetaFlop = 1/3 rack
New Computing Architecture for Data-centric Environments Non Von Neumann Architectures New Architecture & Programming Model New Interconnect, Signaling & Encoding N N Ak New Components & Chips Presynaptic Postsynaptic New Switch & Devices
Data-centric model for computer structure Cognitive Computing
Computers and the Brain are Dramatically Different Separates memory and processor Sequential, centralized processing Ever increasing clock rates, high active power Huge passive power Programmed system, hard-wired, fault-prone Algorithms and analytics Integrates memory and processor Parallel, distributed processing Event-driven, low active power Does nothing better, low passive power Learning system, reconfigurable, fault-tolerant Substrate and pattern recognition
Von Neumann Computing: Left Brain Computing Cognitive Computing: Right Brain Computing Sequential, Analytical Text, numbers, symbolic Front-end, back-end intelligence Parallel, Synthetic Sense-act, sub-symbolic In-situ, physical intelligence Centralized Clock Distributed Event-driven Bus Logical communication by messages No Bus Local physical, global logical communication Cache Memory-inefficient for real-time Registers Overwritten No Cache Fundamentally memoryefficient Update when state changes Programming Hard-wired Fault-prone Algorithms Learning Reconfigurable Fault-tolerant Variable Precision Computation-driven so power-agnostic; Flops Energy-driven so poweraware; Flops / W; scales with activity
Example Program: SyNAPSE Multi-institutional, Multi-disciplinary, Vertically-integrated Approach Potential Applications: - Spatial navigation - Machine vision - Complex System Modeling/Analysis - Pattern recognition - Associative memory - Security $41M in funding from DARPA
Other Approaches to Cognitive Computing Artificial Neural Networks Attributes: Data-Centric New programming model Event-driven, low-power operatioons Integrated memory, computation, & communication Challenges Complexity, New architectures New Programming Model Use of existing CMOS structures & devices inefficient Integration with existing computing structures
Thank you.