Dr. Bernard Meyerson y – IBM Fellow,, Vice President - Innovation,, IBM Design g Automation Conference 2010: Echoes of DACs Past: From Prediction to Realization and Watts Next? Realization, © 2010 IBM Corporation First,, a trip p down memory y lane - 2 Dr. Bernard S. Meyerson © 2010 IBM Corporation The Core Issue Then and Now; Power As Classical Scaling g Ends,, Innovation Accelerates CPU 1 L2 Cache Multi-Core “Lower-Power” Multicore 14 CMOS 12 Bipolar Power4 10 8 6 Junction Transistor 4 Integrated Circuit Water Cooling ? 2 Vacuum tube 0 1950 IBM 360 1960 1970 1980 1990 Year of Announcement 3 SoC / eDRAM Dr. Bernard S. Meyerson 2000 Inno ovation R Required d Module Heat Flu M ux (wattts/cm2) CPU 2 2010 3D Integration © 2010 IBM Corporation Bets ets Placed aced Five e Years ea s Ago……….… go Innovation drives future technology performance – Blind scaling runs one (ran some) into thermal walls Processor design and metrics have changed irrevocably – Gigahertz are gone, replaced by application specific metrics System solutions optimized via Holistic Design will ultimately dominate progress in information technology – Application specific computing revolutionizes computation costs Fiscal reality has driven and is driving our industry towards technological consolidation – Globalization of technology ecosystems is a result 4 Dr. Bernard S. Meyerson © 2010 IBM Corporation Holistic Design; Innovation at All Levels of Systems Stack 5 Clouds Help improve Data Center efficiency System Management Allow better asset utilization Hybrid Systems Enhance emerging workload consolidation S Security it Provide P id ttrusted t d computing ti 10G Ethernet Enable network fabric convergence 3D Integration Revolutionize chip packaging Fl h Flash Increase system performance and efficiency Multi-core Processors Change hardware system design Computational Lithography Dr. Bernard S. Meyerson Extend CMOS density scaling © 2010 IBM Corporation Foundational Innovation; Materials Elements Employed in Silicon Technology Before 90’s 90 s 90’s through 2005 Beyond 2006 6 Dr. Bernard S. Meyerson © 2010 IBM Corporation More of Moore? Not Quite (Don’t Get Comfortable) The Drive for Density Continues, but NOT Linearly Non-planar technology (FINFET) emerges 0.16 μm P PU PD Smallest SRAM cell shown with optical lithography – but based on FINFETs – 0.063 0 063 μm2 cell with 80 nm gate pitch – Cell show excellent functionality down to 0.4 V Vdd – 40 nm fin pitch demonstrated for first time using optical lithography – IBM, GlobalFoundries, Toshiba, and NEC result 0.394 μm SRAM M cell size (μ m 2) 0.4 ITRS Conference report SRAM post fin etch 0.3 This work:0.063 μm2 (SIT-FinFET by ArF) 0.2 40nm fin pitch 0.1 0 45 35 25 15 Tech. Node (nm) 7 Dr. Bernard S. Meyerson 5 TEM of fins © 2010 IBM Corporation Meanwhile-The Search Goes on for A New Switch; One Example - A Graphene RF Transistor Single Layer Graphene Gate length = 200 nm Gate length = 150nm fT = 10 GHz fT ~ 26 GHz measured in a 150-nm-gate graphene transistor The highest fT reported for graphene so far 8 Dr. Bernard S. Meyerson Lin et al., Nano Letters 9, 422-426 (2009) © 2010 IBM Corporation Electrical Bandwidth Bottlenecks Emerge Forcing Optics Innovation and Integration • Electrical Buses become increasingly difficult at high data rates (physics): • 10000 Optical Backplane: 10 GB/s, 62.5μm pitch Increasing losses & cross-talk ; Frequency resonant affects Estimated Limit of MCM Electrical Escape BW Optical data transmission: • Power Efficiency , much less lossy, not plagued by resonant effects BW W (GB/s) • Off-card Bandwidth, GB/s Limit of Electrical Backplane BW 1000 2015 2011 100 10 2 Core 2 Core 2 Core 2 Core 2 Core 4 Core 8 Core 16 Core 32 Core System or Time Rack p Backplane Card Module μP OPTICS μP MULTI CHIP MODULE CIRCUIT BOARD 9 Dr. Bernard S. Meyerson © 2010 IBM Corporation Packaging Innovation Enabling 3D Integration of Logic, Memory and Optics 3D Integration allows restructuring of the compute node to leverage dense memory and dramatically increase memory bandwidth This produces significant performance improvements with necessary software coevolution/adaptation p Optics: Multi-cores Multi cores require huge bandwidth, bandwidth but the cost of optics is dropping rapidly, allowing Optics to move inside the system: rack to card to chip ptical tr affic al I/O On-chip o Off-ch ip Optiic optica ls ignals CMOS integrated silicon photonics can provide large power savings for on-chip interconnects Photonic Plane Memory Plane Logic Plane Logic plane Memory plane Photonic plane 10 Dr. Bernard S. Meyerson ~300 cores ~30GB eDRAM On-Chip Optical Network >1Tbps optical on-chip >1Tbps optical off-chip © 2010 IBM Corporation I Could Follow This Thread For Several M More Hours, H B Butt That Th t Would W ld Be; B DEATH BY POWERPOINT So Instead, Visit One’s CFO to Explain… © 2010 IBM Corporation The Impact of Physics on Finances DEATH BY CFO © 2010 IBM Corporation All This Innovation Did Not Come Cheap! Extraordinarily complex and costly solutions have emerged to address; –Lithography (Multiple exposure, EUV,……..) –Transistors (FinFet’s, Graphene, ……) –Interconnects Interconnects (3D/Stacking (3D/Stacking, Optical Optical,…)) –Manufacturing (32nm->22nm fabs) EUV at Albany Nanotech 13 Dr. Bernard S. Meyerson © 2010 IBM Corporation Was/Is There a Sustainable Business Model? Chip Making R&D Versus Revenues 6 10 The industry trend as of today: R&D Expenses p far outpace p revenues 2004-2020 Projected CAGR With VLSI inc. (Worldwide in Permission, $M) ~$2.3B 5 10 Estimated cost to develop the 4 1022-nm CMOS logic process 1,800 3 10 2 1,400 , R&D ~ 12 12.2% 2% $1.1B 10 Process development Revenue Growth 6.5% 2004 - 2010 CAGR This was NOT sustainable 1,000 Process ramp-up 1 10 0 800 10 70 600 60 400 50 Revenu ue/RD&E R&D Expenses 12.2% Costs ($M) C 1,200 Total RD & E (Chip + Eq) Semiconductor Revenues extrapolate ed 1,600 Revenues ~ 6.5% 200 40 Fiscal reality drove our industry to 0 consolidate around Innovation 30 180nm 130nm 90nm 65nm 45nm 32nm 22nm Networks. Source: IBS Global 20 System IC Service Management Report, April 2006. The predicted consolidation continues virtually unabated. unabated 14 Dr. Bernard Meyerson Dr.S.Bernard S. Meyerson 10 0 1960 1970 1980 1990 2000 2010 2020 Year Source: VLSI Research inc. © 2008 IBM Corporation Semiconductor Technology Globalization Intel AMAT GF Freescale SEMATECH R&H IBM IMEC Infineon GF STM TNS Hitachi TEL NEC ToshibaSamsung Sony JV Toppan SELETE TSMC GF(Chartered) Equipment and Materials Partners with IBM IBM Partners CEA-LETI Partners Major Consortia-IMEC&SEMATECH 15 Dr. Bernard S. Meyerson © 2010 IBM Corporation Emergent-Silicon Valley East; The Hudson Valley Emergent2005--2008 2005 Yorktown Research Filtering New Materials & Processes Albany R&D 32 nm node and beyond Process & Tool R&D Process Window / Manufacturing Scaling / Learning Rates Early Defective & Yield Assessment New metrology/characterization methods Production Scaling-Fishkill IBM Partners 16 Dr. Bernard S. Meyerson Materials & Equipment Suppliers Matheson Tri Gas © 2010 IBM Corporation Albany NanoTech Evolves to be a Full Flow Fabricator 2005 Panoramic and aerial views of the College of Nanoscale Science and Engineering (CNSE) facilities, which houses the Center for Semiconductor Research (CSR) Line where IBM is performing 22nm advanced semiconductor research and development 2009 NanoFab South Annex (NFX) NanoFab North (NFN) NanoFab N F b East (NFE) NanoFab S th South (NFS) 17 Dr. Bernard S. Meyerson © 2010 IBM Corporation An Ecosystem Grows Based Upon Radical Collaboration 2009--………. 2009 Design IP Advanced Research Design Tools Manufacturing Platform NYS Partnerships CCIC Manufacturing Design Services Design Tools Ecosystem SOI & Bulk CMOS Joint Process Development Alliances Tools & Materials IP Solutions Computational Scaling Matheson Exploratory Research Library Solutions Packaging Joint Development De elopment NRI 18 Dr. Bernard S. Meyerson © 2010 IBM Corporation The New Paradigm; Radical Collaboration as Taught at HBS Science-Based Business and the Business of Science 19 Dr. Bernard S. Meyerson © 2010 IBM Corporation Acknowledging A Discontinuity; Information Technology gy Reinvents Metrics © 2010 IBM Corporation Systems and Technology Group A Prime IBM Example Of How Metrics Have Evolved; The Clock Race Terminates Where have all the “Giga Hertz” gone? 21 Dr. Bernard S. Meyerson © 2010 IBM Corporation New Metrics for Processors •Compute Compute capacity in a given System implementation •Application Adaptive Multicore designs y y address throughput g p and/or thread p performance apps pp •Dynamically •Autonomic energy efficiency & management capabilities •Advanced enablement of platform and systems management •Virtual Server Placement •Image Management •Storage/Network Integration The Bottom Line;; We Must Now Focus On Solutions <<< Solution >>> process of solving gap problem The action or p 22 Dr. Bernard S. Meyerson © 2010 IBM Corporation Application Specific Problem Solving” Solving “Problem Stream Computing: A paradigm di for f handling h dli the flood of data © 2010 IBM Corporation What Flood? By 2011, the world will store 10X the Data stored in 2006, BUT; 1,800 1,600 10x growth th in i Data D t over five years Exabytes s 1,400 RFID RFID, 1 200 1,200 Digital TV, 1,000 MP3 players, Digital cameras, 800 Camera phones, VoIP, Medical imaging, Laptops, 600 smart meters, multi-player games, 400 Satellite images, GPS, ATMs, Scanners, Sensors Digital radio, Sensors, radio DLP theaters, theaters Telematics , Peer-to-peer, Email, Instant messaging, Videoconferencing, CAD/CAM, Toys, Industrial machines, Security systems, Appliances 200 0 2005 2006 2007 2008 2009 2010 2011 I Internet connected d devices d i will ill grow by b 2000X, 2000X from f 500M to 1 T Trillion, illi and each will demand that someone “listen”. 24 Dr. Bernard S. Meyerson © 2010 IBM Corporation An Application Specific Solution Stream Computing, aka How not to drown in data A new paradigm for ultra low latency and high throughput in-motion in motion analytics Continuous Ingestion 25 Dr. Bernard S. Meyerson Continuous Queries /Analytics on data in motion © 2010 IBM Corporation Streams Processing Building Blocks Transform NYSE Dynamic P/E Ratio Calculation Milllions of Eve ents VWAP Calculation SEC Edgar 10 Q Earnings Extraction Caption Caption Extraction Caption Extraction Extraction Video News Video News Video News Weather Data Speech Speech Recognition ecog to Speech Recognition Recognition Hurricane Weather Data Extraction Filter Annotate Correlate Join P/E with Aggregate Impact Earnings Moving Average Calculation Topic Topic Filtration Topic Filtration Filtration Earnings Earnings Related Earnings Related News Related News Analysis News Analysis Anal sis Analysis Hurricane Forecast Hurricane Model 1 Forecast Hurricane Model 2 Forecast Hurricane Model … Forecast Model N Earnings News Join Hurricane Risk Encoder Trade Decision Hurricane Impact Hurricane Industry Impact Classify < 1 millisecond latency 26 Dr. Bernard S. Meyerson © 2010 IBM Corporation TD Bank Building a next generation trading platform Identify and execute trades Process over 5M events per second with average latency of 150 microseconds Expand to incorporate content feeds, news text, audio, video, to establish greater context for better decisions CIO TD Bank B k "TD Bank B k Financial Fi i l Group G worked k d with ith IBM R Research h tto d develop l a first-of-a-kind fi t f ki d architecture capable of consuming, analyzing and acting on real-time market data while maintaining sub-millisecond response times even under extreme data loads” 27 Dr. Bernard S. Meyerson © 2010 IBM Corporation Life-Changing When Applied in Medicine University y of Ontario Institute of Technology gy A Research Project to monitor premature infants in the ICU Correlating blood oxygenation with blood pressure to predict “Baby crashing” Nosocomial N i l IInfection f ti Prediction –Monitoring heart rate variability with other information to predict sepsis –Alarms Alarms up to 24 hours earlier than by experienced ICU Nurses Sources On-the-fly stream computing InfoSphere Streams Files, TCP Sockets Persist/Enrich solidDB Persist stream data ODBC/JDBC/SA solidDB, DB2, IDS InfoSphere Warehouse http://www.youtube.com/ibmhealthcare // / 28 Dr. Bernard S. Meyerson © 2010 IBM Corporation Cloud Computing; Solutions on a Global Scale © 2010 IBM Corporation IBM Systems and Technology Group Why Not Stay With What We Know? Spending (US$B) Installed Base (M Units) Worldwide IT Spending Trend $300 $250 Power and cooling costs 50 Server mgmt and admin costs 45 New server spending 40 35 $200 30 $150 25 20 $100 15 10 $50 5 $0 10 08 09 20 20 20 07 20 06 20 05 20 03 04 20 02 00 01 20 20 20 20 98 99 19 19 97 19 19 96 0 We cannot continue our past strategy; Server management costs dwarf all others Datacenter power demands are growing at unsustainable rates – 1.2% of global electrical output is used by servers and related cooling equipment – Spending on datacenter power is growing 600 600-800% 800% faster than spending on servers. Source: IDC, Virtualization 2.0: The Next Phase in Customer Adoption, Doc #204904, Dec 2006 30 Dr. Bernard S. Meyerson © 2010 IBM Corporation Economy y of Scale is Driving g IT From Left to Right g Complex p Physical Consolidation Logical Simplification Service Managed Data Centers Windows Servers Windows Server Firewalls, Unix Servers Virtual Servers, Storage, Networks Assets Routers Unix Server Switches Networks Virtualization Management Assets Assets Servers Linux Server Linux Servers Storage Storage Servers Storage Assets Networks Build It Yourself Requirements • Lower management costs • Reduced time to deploy applications, including self-service • Higher availability, security, energy efficiency, resource utilization • Service managed infrastructure 31 Dr. Bernard S. Meyerson Grouping of like systems Complexity hiding Self-service to resources, applications Request driven provisioning Automated application on-boarding Appliances © 2010 IBM Corporation Applying Cloud Computing to the EDA Industry 32 Dr. Bernard S. Meyerson © 2010 IBM Corporation Exponential Growth of Transistors But Not of Designers W Must We M t Make M k This Thi Process P “Easier” “E i ” 1.8 10,000 1.4 1.2 1 1,000 0.8 0.6 0.4 Power uP Transisto P ors M Designers (u uP Design) rel. units 1.6 Designers Designers/Transistor Transistors 0.2 Dr. Bernard S. Meyerson 11 20 10 20 09 20 08 20 07 20 06 20 05 04 20 20 03 20 02 20 01 20 00 20 19 19 33 99 100 98 0 © 2010 IBM Corporation Compounding the Challenge, Design Issues are Ever More Complex, Even in This Back-level Back level Assessment! Functionality + Testability Thousands Functionality + Testability + Wire Delay F Functionality ti lit + Testability T t bilit + Wire Wi Delay D l + Power P Mgmt M t # Transistors Functionality + Testability + Wire Delay + Power Mgmt + Embedded Software Functionality + Testability + Wire Delay + Power Mgmt + Embedded Software + g Integrity g y Signal Functionality + Testability + Wire Delay + Power Mgmt + Embedded Software + Signal Integrity + Hybrid Chips Functionality + Testability + Wire Delay + Power Mgmt + Embedded Software + Signal Integrity + Hybrid Chips + RF Functionality + Testability + Wire Delay + Power Mgmt + Embedded Software + Signal Integrity + Hybrid Chips + RF + Packaging Billions Functionality + Testability + Wire Delay + Power Mgmt + Embedded Software + Signal Integrity + Hybrid Chips + RF + Packaging + Mgmt of Physical Limits •Exponentially growing number of elements •Design D i complexity l it is i exponential ti l function f ti off element l t countt Source: Semiconductor Research Corporation 34 Dr. Bernard S. Meyerson © 2010 IBM Corporation Cloud Computing Solutions Are Highly Tuned To Workloads – Flexible, Fl ibl configurable fi bl cloud l d off virtualized i t li d iimages 1-1000 Servers • Ease of management and cost reduction of many different system images. *** *** • W Workloads kl d & A Application li ti D Development l t&T Testt systems Cloud Burst Many Different Images • Dynamic, user configurable systems – e.g. Amazon EC2,, IBM Cloud Burst,……. , – Cloud of highly utilized common systems for batch intensive workloads and priority interactive jobs • Ease of management and cost reduction of compute intensive workloads. • Workloads & Applications Batch jobs of Engineering applications Interactive jobs with same image • Maniacally eliminate computational overhead to increase utilization. • Lower cost per node 35 Dr. Bernard S. Meyerson 10-100,000 Servers *** Few Images *** Design Cloud © 2010 IBM Corporation IBM Design Cloud History and Roadmap IBM has been using a design cloud/grid for more than 20 years. – Begun in the late ‘80s as a grid of servers and PC’s doing logic simulation/verification. – In 2Q of 2009 all designers were using the design cloud. Today – All IBM chips are designed in the cloud – IBM Design Cloud • 20,000+ Cores, 150+ TB RAM, 1+ PB Disk in production across 3000+ Users • 40K+ Jobs/day, 50M+ Sim cycles (processor clocks) – Global EDA Cloud infrastructure and user community – Tool sets; • Logic design, Circuit design, Layout, Verification, IBM EDA tools, Vendor tools 22nm designs have launched on the Design Cloud. C – Infrastructure upgraded to handle larger chips/transistor counts. – Collaboration with Mentor to minimize OPC runtimes. • 32nm OPC uses a Roadrunner Supercomputer. 36 Dr. Bernard S. Meyerson © 2010 IBM Corporation Design Cloud Delivery Management System Batch (90%) Interactive (10%) Optimized for high-utilization high utilization through a balance of batch and realtime jobs Cloud Achieves 80-90% CPU utilization / 60% memory utilization. g Cloud was enabled by y the ability y to support pp remote high g The Global Design resolution graphics Interactive user jobs run at maximum priority for designer efficiency –Over 3000 demanding Design Cloud users •Interactive I t ti jobs j b are performance f critical iti l tto maintain i t i user acceptance. t Unique computing Resources are deployed for maximum efficiency –OPC mask preparation requires a high-performance supercomputer –Logic simulation for virtual Bring-up and Power-on requires a massive FPGA system. –Must support varied computer architectures and operating systems •Today – Power, x86, SUN, Roadrunner and AWAN(FPGA) 37 Dr. Bernard S. Meyerson © 2010 IBM Corporation Data Wins; High Utilization is Consistently Achieved High utilization (~ 90% steady state) for 48 hours while handling both batch and interactive jobs simultaneously 38 Dr. Bernard S. Meyerson © 2010 IBM Corporation The (New) Bottom Line • Innovation Drives Performance, Period. • There are no magic bullets left, just a lot of hard work. • System, not chip performance, is the new metric • A focus on enabling Solutions is emerging; e.g. Stream p g, Application-Specific pp p Acceleration,, Appliances,……… pp , Computing, • Cloud computing is emergent as a class of Solution as applied to EDA • EDA Clouds have proven track records in supporting demanding users on a global scale. • Economy of scale, agile and adaptive capacity, and the conversion from Capex to Opex all favor this trend. • The transformation of this industry is still accelerating 39 Dr. Bernard S. Meyerson © 2010 IBM Corporation Backup © 2010 IBM Corporation Scaling streams using cloud computing - from a single instance to thousands VWAP Timeperiod NYSE Calculate P/E Ratio RSS Parser Keyword Filter Deployment on on the Deployment High Performance Cluster Cloud News Join P/E and Aggregate Impact Semi Annual Moving Average SEC Edgar Keyword Filter Gister RSS Parser RSS Parser Keyword Filter Hurricane Forecast Hurricane Impact Hurricane Industry Impact Weather Data VWAP Timeperiod NYSE Calculate P/E Ratio RSS Parser Keyword Filter News Join P/E and Aggregate Impact Semi Annual Moving Average SEC Edgar Keyword Filter Gister RSS Parser RSS Parser Keyword Filter Hurricane Forecast Hurricane Impact Trade Decision Hurricane Industry Impact Weather Data VWAP Timeperiod NYSE Calculate P/E Ratio RSS Parser Keyword Filter Semi Annual Moving Average SEC Edgar News Keyword Filter Gister RSS Parser RSS Parser Keyword Filter Hurricane u ca e Forecast Hurricane u ca e Impact Join P/E and Aggregate Impact Hurricane Industry Impact Weather Data 41 Dr. Bernard S. Meyerson © 2010 IBM Corporation The Cloud over the years Network-Delivered Network Delivered Services are the culmination of a long term trend to simplify the purchase of IT Services 1961: John McCarthy proposes computing as a utility 1961: IBM Services Bureau 1975: First inter industry EDI standards inter-industry 1981: SMTP defines the standard electronic mail service 1985: United Nations sponsors EDIFACT 1990: Berners-Lee invents the World-Wide Web 1994: CommerceNet 1998: RosettaNet 1999: i-Mode mobile internet 2000: IBM BCRS 2000: UDDI 1.0; “SaaS” SaaS coined 2001: Dot com bubble bursts IBM Service Bureau (1961) 2005: IBM AoD 2006: Amazon EC2 2007: Google Health; force.com launch 2008: IBM ww Cloud Computing centers 19601970 42 1980 Dr. Bernard S. Meyerson 1990 2000 2010 © 2010 IBM Corporation Innovation Across Novel Materials, Structures, Processes, and d Architecture A hit t Continues C ti … Airgap eDRAM 3D Chip Stacking High-k Frozen SiGe Chip Immersion Dual Core SOI Strained Silicon Slowing Speed p of Self Light Assembly Copper Atomic p Manipulation Nanotube IC Atomic Molecular Storage Processing Nanophotonic S itch Switch Highest Resolution Carbon Nanotube STEM Transistors Nobel Prize, STM 43 Dr. Bernard S. Meyerson © 2010 IBM Corporation Application Specific Computing Within a Cloud More ways to skin a compute intensive cat; System X (dx360 M3) – Server w/embedded GPU Configuration Announced – May 18, 2010 4- 2.5” SS SAS 6Gbps (or SATA, or 3.5”, or SSD…) NVIDIA Tesla M2050 #2 NVIDIA Tesla M2050 #1 (or NVIDIA Quadro FX3800, or Fusion IO, or…) ( NVIDIA T (or Tesla l M1060 M1060,or FX3800 FX3800, or Fusion F i IO, IO or)) Server level value 44 Each server ser er is individually indi id all serviceable ser iceable Each GPU is individually replaceable 6Gbps SAS drives and controller for maximum performance Service and support for server and GPU from IBM Dr. Bernard S. Meyerson Infiniband DDR (or QDR, or 10GbE…) © 2010 IBM Corporation Design Cloud Criteria for Success Extremely high server utilization –minimize cost while maintaining high performance globally Simplified maintenance and scalability –Servers and storage co-located for ease of scaling. –Revision Revision control readily effected with no shadowing to remote locations. High security, availability and reliability –Sony, Microsoft and Nintendo designs were done in the Design Cloud. –When When a servers goes down all batch jobs are restarted restarted. –Interactive jobs do not lose data given uniform backup policies. Efficient use of design licenses. –Designers from around the world use the licenses on a pool of servers. High resolution graphics must be accessible globally. –Improves p designer g p productivity. y 45 Dr. Bernard S. Meyerson © 2010 IBM Corporation