IBM Research: Science & Technology and Semiconductor Technology Research Overview

Home , GlobalFoundries, IBM (atoms), IBM Quantum Experience, IBM Research

IBM Research

§ IBM Research Division - quick overview

§ Semiconductor Technology Research (STR) – Business Model – Roadmap – Collaboration

§ Science & Technology (S&T) – Mission – Infrastructure – Research Programs – Collaboration

Science & Solution Twelve Labs with Over 3,000 Researchers Around The World Strategy Labs Labs Opened Since 2011

Tokyo (2012/1982) Zurich (1956) Almaden (1986/1952) Watson (1961) Shin-Kawasaki Rueschlikon, San Jose, CA Yorktown Heights, NY & Toyosu, Japan Switzerland

First New Research Lab in 12 Years Dublin (2011) China (1995) AustinBrazil (2010)(1995) Dublin, Ireland Beijing, China Sao PauloAustin, &Rio TXde Janeiro Shanghai (2008)

Brazil (2011) Africa (2012, 2015) Haifa (1972) India (1998) Australia (2011) Sao Paulo & Nairobi, Kenya & Haifa, Israel Delhi, India Melbourne, Victoria Rio de Janeiro Johannesburg, S.A.

30 November, 2016 3 A Diversity of Disciplines From Atoms to Service Science

Physics Chemistry Materials Science Electrical Engineering

Computer Sci. Mathematical Sci. Behavioral Sci. Service Science

30 November, 2016 4 2016 Quantum Computing in the Cloud A Legacy of World-Class Research 2015 Quantum Computing: Error Detection For 70 Years 2015 23rd Consecutive Year of Patent Leadership 2012 Atomic Imaging (Charge Distribution, Bond Order) 2011 Watson System 2009 Nanoscale Magnetic Resonance Imaging (MRI) 2008 World’s First Petaflop Supercomputer 2007 Web-scale Mining 2005 Cell Broadband Engine 2004 Blue Gene/L 2003 5 Stage Carbon Nanotube Ring Oscillator 2000 Java Performance 1998 Silicon on Insulator (SOI) 1997 Copper Interconnect Wiring 1997 Deep Blue 1994 Design Patterns 1994 Silicon Germanium (SiGe) 1990 Chemically Amplified Photoresists 1987 High-Temperature Superconductivity (Nobel Prize) 1986 Scanning Tunneling Microscope (Nobel Prize) 1980 Reduced Instruction Set Computing (RISC) 1979 Thin Film Recording Heads 1973 Winchester Disk Drive 1971 Speech Recognition 1970 Relational Database 1967 Fractals 1966 One-Device Memory Cell 1957 FORTRAN 1956 Random Access Memory Accounting Machine (RAMAC)

30 November, 2016 5 A Culture of Innovation – External Recognition 1 Millennium 2 Kavli Prizes 6 Turing Awards 6 Nobel Laureates Technology Prize

10 US 2 U.S. 5 US National 1 Emmy Presidential National Medals of Medals of Medals of Technology Freedom Science

22 Members in National 64 Members in National > 400 Professional 20 Inductees in National Academy of Sciences Academy of Engineering Society Fellows Inventors Hall of Fame

v AAAS v APS v IEEE v ACM v AVS v IOP v ACS v ECS v OSA

30 November, 2016 6 Innovations That Matter Have Enabled an Unprecedented 23 Consecutive Years of IBM Patent Leadership

• 23 Consecutive Years of Patent Leadership 8,000 • It took IBM 50 years to receive the first 5,000 patents and now routinely produces over 5,000 per year 7,534 7355 • IBM's first patent: U.S. Patent #998,631 issued July 25, 1911. 6,809 7,000 6,478 6,180 5,896 6,000 4,914 5,000 4,186 4,000 3,411 3,415 3,621 3,288 2,886 3,248 3,125 2,756 2,941 3,000 2,658 1,867 2,000 1,724 1,2981,383 1,085 1,000 0 2011 2011 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2012 2013 2014 2015

30 November, 2016 7 Evolution of IBM Research Business Model Business Model Evolves as Business Conditions Change In Order for IBM Research to Remain Essential to the IBM Company

1970's 1980's 1990's 2000's 2010's § Corporate § Collaborating § Work on § Industry-focused § Collaborative funded across IBM client research Partnerships research problems § Shared § Created Bus. § Emerging agenda agenda Adv. For Clients markets § Technology § Effectiveness Globalize & Engage transfer Client Business Research Research in the Marketplace Joint Programs Centrally Funded

30 November, 2016 8 IBM Research: Aligned to Growth

Cloud AI Data + AI Transforming Transforming Transforming IT Computing Industries

Computing Cognitive Cognitive as a Service Platform Solutions

Science and Emerging Technologies

30 November, 2016 [email protected] | IBM Conﬁdential | For Use Only by High Level Steering Committee (HLSC) of the EU Quantum Technology Flagship Project 9 IBM Research Semiconductor Technology Research IBM Relationship with GlobalFoundries Acquisition of IBM’s Semiconductor Business by GLOBALFOUNDRIES is Completed Armonk, N.Y. - 01 Jul 2015: IBM (NYSE: IBM) today announced that the acquisition of the company’s global commercial semiconductor technology business by GLOBALFOUNDRIES has been completed. With the closing of this transaction, GLOBALFOUNDRIES becomes IBM's exclusive semiconductor processor technology provider for the next 10 years, ensuring a long-term supply of semiconductors for IBM systems. IBM Research will continue its deep focus on fundamental semiconductor and material science research and systems innovation to drive IBM’s leadership in mainframe, Power and storage systems as well as future cloud, big data and analytics systems. “This announcement is the next step in our long-standing relationship with GLOBALFOUNDRIES. IBM continues to invest in systems leadership, innovation and talent for the long-term,” said Tom Rosamilia, senior vice president, IBM Systems. “IBM is designing and developing IT systems for the digital era -- including servers, storage and middleware that will empower our clients to drive new workloads and new business models.“ IBM expects its unparalleled semiconductor and materials research, which has delivered such breakthroughs as copper chips, silicon germanium and quantum computing research, to continue to advance its capabilities in systems for years to come.

IBM/GF Relationship Is Imperative to Future IBM Systems GF Must Successfully Manufacture Microprocessors for IBM Mainframes / Systems

…let me be clear-we are not exiting hardware. IBM will remain a leader in high-performance and high-end systems, storage and cognitive computing, and we will continue to invest in R&D for advanced semiconductor technology. Cloud and big data applications are placing new challenges on systems, just as the underlying chip technology is facing significant physical scaling limits. Virginia Rometty, IBM CEO IBM 2013 Annual Report IBM News Room March 8, 2014 July 10, 2014

Partners with GLOBALFOUNDRIES, Samsung and SUNY Polytech to Clear Path

For Next Generation Semiconductors

IBM News Room July 9, 2015

Mission: Industry Leadership in Semiconductor Technology Research

Almaden, CA USA Haifa, Israel Tokyo, Japan Lithography Materials mmWave Technology Neuromorphic Sciences Exploratory Physics THz passive imaging Nano/micro systems for low power Integrated sensors IoT device minaturization Intelligent Machines Optical Interconnects

Yorktown, NY USA Albany, NY USA Zurich, Switzerland Quantum Computing Center for Semiconductor Research Binnig & Rohrer Nanotechnology Center Neuromorphic Computing 10nm / 7nm / Pathfinding III-V & Si Photonics MRAM & PCM EUV Center of Excellence Low Power technologies III-V & Si Photonics New device architectures Quantum Technologies Wearables III-V, 3DI, MRAM Precision Diagnostics

12 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Industry 1st : IBM Semiconductor Technology Research 7nm Device Demonstration @ Albany with Early R&D Partners § Key Features: – Fully Integrated build – Sub 50nm gate pitch, 15nm Lpoly – Sub 30nm fin pitch, dual channel – Sub 40nm BEOL pitch – Multiple EUV levels Sub 30nm pitch, dual channel fins Sub 50nm gate pitch w/ SA contacts Will IBM’s Ultra- Dense Chip Be A Game Changer?

13 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Semiconductor Technology Research IBM Semiconductor Technology Business Model

§ IBM is focusing on Industry Leadership in Semiconductor Technology Research, and will transfer the results to manufacturing partner with a leading edge, at-scale fab

IBM Research IBM & Partners Foundry Partner Science & Semiconductor Technology Technology Technology Development & Research Research Manufacturing Almaden, Yorktown, Albany Nanotech Tokyo, Zurich & Haifa Center East Fishkill / Malta

IBM Systems Semiconductor Design and Product Engineering

NET: Path for IBM Research Technologies to Market Through Manufacturing Foundry Partner

14 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Semiconductor Technology Research New Device IBM Semiconductor Technology Roadmap 5nm Architecture 3rd Gen FinFET 7nm w/ EUV & new channel

Generations of 10nm Scaled FinFET Technology Leadership

FinFET 14nm 3rd Gen. HiK R&D à à 2nd Gen. HiK 22 nm Production

High-K gate 32 nm dielectric on SOI 45 nm eDRAM, Immersion lithography Ultra Low-k BEOL insulators

65 nm Advanced Strained Silicon

90 nm Low-k dielectrics + DSL

130 nm 2nd Gen. H/S PDSOI

180 nm H/S Silicon-on-Insulator 15 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Semiconductor Technology Research Scaling vs. Innovation Driving the innovation pipeline is critical as the benefits from traditional scaling decline Semi-conductor materials and device innovaon 100%

90% 90nm and beyond: Gain by Innovation 80% Strained Silicon 70% 32nm and beyond: High K / Metal Gate 60% 14nm and beyond: 50% FinFET New Device 40% Architecture 30% Gate Fin 20% Traditional Gain 200 nm 10%

Relative Performance Improvement Relative Performance 0% 180nm 130nm 90nm 65nm 45nm 32nm 22nm 14nm 10nm 7nm 5nm Gain by Innovation, Materials, and Structures Traditional Gain Prior to 90nm: performance improvement was from scaling … Now: innovation, materials and structures are key 16 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Semiconductor Roadmap: Technology Innovations Continue Performance Gains High-k / Si Metal Gate SMO Mask Design Nano- Extending Si CMOS wires Interconnects / Strain Layers ULK Dielectric FINFETs MRAM, PCM Silicon 3Di for SCM Photonics Subsystem Integration eDRAM

III/V Non-Si FET Q2 Metal Gate high-κ Q1 25 nm 45 nm Q3

Beyond Von- Beyond FET Neumann Carbon Architectures Electronics

§ Semiconductor Technology Research (STR) – Business Model – Roadmap – Collaboration

§ Science & Technology (S&T) – Mission – Infrastructure – Research Programs – Collaboration

18 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Science & Technology Science & Technology (S&T) is the Foundation for IBM Research

Data Engaging Cloud Transforming in New Transforming Industries & Ways IT Government

Industries & Cloud Cognitive Solutions Research Research Research

Science & Technology Research

Provide IBM Systems & Cloud with Create new growth opportunities competitive advantages delivered for IBM and our partners by through leadership in materials advancing the frontiers of and processes, packaging, logic information technology through and memory devices, I/O and the physical sciences and computation subsystems hardware innovations

Yorktown, NY Almaden, CA Zurich Tokyo Haifa USA USA Switzerland Japan Israel

Providing the Ability to Rapidly and Directly Determine Structures of New Molecules Not Determinable by Other Analysis Methods § Molecules: Metabolite from Dermacoccus abyssi collected in the Mariana Trench at 10898 m depth – Composition determined from mass spectrometry: C16H10N2O2 – Structure not known (x-ray crystallography not possible due to lack of material) +2 Hz O O N f scale f Δ N H

-7 Hz AFM Measurement, AFM Measurement, Constant Height, 3D Molecular Model Proposed Molecular Structure: Representation Overlaid Cephalandole A L. Gross et al., Nature Chem. 2(10), 821 (2010) 21 July 2015 [email protected] | IBM Confidential © 2015 IBM Corporation IBM Research Science & Technology Microelectronics Research Laboratory (MRL)

MRL offers state of the art processing and the unique opportunity to utilize existing CMOS technology and further integrate novel processes to build nanofabrication capabilities not available anywhere else

Poly-Si SiO 2

Hf-based dielectric & 5 nm TaN

Silicon nanowires, Photonics, 3D Non volatile Nanochannel New devices and III-V devices integration (TSV) memory (PCM, electrodes for nanopattering: Directed MRAM)-integration sensing self assembly, CNT, A sample of MRL Capabilities… of new materials biomolecules QUBIT 22 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Science & Technology Science & Technology Strategic Imperatives § Continue silicon scaling

10 July 2014

§ New materials and devices to extend core Scaling: 22, 14, 10, Gate All Around III/V logic, memory & I/O technology roadmaps 7, 5 nm Nodes Devices Devices § New computing devices and architectures

PCM 3D

Silicon Carbon Photonics Devices MRAM

Quantum Cognitive Computing Computing

Kailash Gopalakrishnan

Security, Fraud Detection 18 Graph Genome Analysis > 10 Analytics Social Network Analytics Knowledge Graphs

Flops 100x Watson Health for Image Classification size increase Watson Analytics Machine in past 5 years Robotics Learning Education > 100 MW 7 > 10 hrs Multimodal Analytics to process Video YouTube videos For Speech Training - Object recognition Analytics from a single day In 1000’s of languages - Complex video analytics - Correlation and stitching

§ Cognitive Workload computational requirements growing significantly. § Saturation in Silicon technology, circuit & architectural performance

Cognitive Application Analysis § Application analysis & profiling is the first step - Deep Learning, Machine Learning, towards understanding bottlenecks in today’s Robotics (optimized) implementations. § Exploitation of commercially available accelerators Exploitation of Current – FPGA, GPU etc – to get ~ 5-20X in performance Accelerators over CPU. - FPGA, GPU, Synapse § Design of new accelerator microarchitectures that Design of New Accelerators for overcome the bottlenecks & programming-model the Cognitive Era challenges of today’s accelerators. § Further 10-100X improvement – exploiting approximations in synchronization, precision & Approximate Architectures perforation over current accelerators. § Extending approximations down to the circuits & Approximate Circuits, device level – analog computing, computing-in- Devices & Memory memory, approximate circuit design & synthesis.

GPU/Other NVLINK Deep Advantages: Learning POWER 8+/9 § 5-20X Improvement

Processors in performance

NVLINK GPU/Other Analytics PowerCore Platforms Challenges: Coherent GPUs with Power9 § Works only for simple applications FPGA Video AFU Analytics & (not fragmented IBM Supplied PSL Coherent FPGAs Transcoding code) with Power8 § Programming Model PCIe Challenges

CAPP Database Acceleration § Cost of Offload

POWER8 Core § SW Debug /

OS ….. Development App

POWER8 Processor

Cloud enabled Cognitive Systems • Novel data ingestion mechanisms Near designed for stochastic sensory Term inputs

Approximate and error tolerant Algorithms • Stochastic Programming Models • Real time algorithms • Compilers Approximate and Stochastic Chip Medium Architectures • Automatic synthesis of Term Approximate and error resilient circuits Approximate and error resilient Circuits • Modeling of Approximate and Stochastic logic and memory Long devices Term Approximate and Stochastic Devices

Approximate7 Computing for Deep Learning

iue3 ogee ewr ihalteblsadwhistles and bells the all with network GoogLeNet 3: Figure

input

7x7+2(S)

Conv

3x3+2(S)

MaxPool

LocalRespNorm 1x1+1(V)

Conv Role of

3x3+1(S) Conv

LocalRespNorm Reduced

3x3+2(S)

MaxPool

3x3+1(S) 1x1+1(S) 1x1+1(S)

MaxPool Conv Conv

1x1+1(S) 5x5+1(S) 3x3+1(S)

1x1+1(S) Precision Conv Conv Conv Conv

DepthConcat

3x3+1(S) 1x1+1(S) 1x1+1(S)

MaxPool Conv Conv

1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S)

Conv Conv Conv Conv

DepthConcat

3x3+2(S)

MaxPool

3x3+1(S) 1x1+1(S) 1x1+1(S)

MaxPool Conv Conv

1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S)

Conv Conv Conv Conv

DepthConcat

5x5+3(V) 3x3+1(S) 1x1+1(S) 1x1+1(S)

AveragePool MaxPool Conv

Conv Perforation

1x1+1(S) 1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S)

Conv Conv Conv Conv Conv

FC DepthConcat

3x3+1(S) 1x1+1(S) 1x1+1(S)

MaxPool Conv Conv

1x1+1(S) 5x5+1(S) 3x3+1(S)

1x1+1(S) Role of Loop

SoftmaxActivation

Conv Conv Conv Conv

softmax0 DepthConcat

3x3+1(S) 1x1+1(S) 1x1+1(S)

MaxPool Conv

Conv Perforation

1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S) Input Maps Kernels Output Maps

Conv Conv Conv Conv

DepthConcat CIFAR10 Benchmark

5x5+3(V) 3x3+1(S) 1x1+1(S) 1x1+1(S)

AveragePool MaxPool Conv Conv

1x1+1(S) 1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S)

Conv Conv Conv Conv Conv

FC DepthConcat

3x3+2(S)

MaxPool

3x3+1(S) 1x1+1(S) 1x1+1(S)

SoftmaxActivation

MaxPool Conv

Conv Power or x86 cluster

1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S)

softmax1

Conv Conv Conv Conv

DepthConcat

3x3+1(S) 1x1+1(S) 1x1+1(S)

MaxPool Conv

Conv Role of Relaxed

1x1+1(S) 5x5+1(S) 3x3+1(S) 1x1+1(S)

Conv Conv Conv Conv

DepthConcat

7x7+1(V) Synchronization AveragePool

SoftmaxActivation softmax2

§ Overall 10-100X Performance gains possible by exploiting approximations across the stack § Robotics, Graph and other Applications currently being studied 6 October 2015 IBM Proprietary - For U.S. Government Use Only - Not for Redistribution © 2015 IBM Corporation IBM Research Going Deeper In the Stack with Approximations… There is plenty of room at the bottom – for exploitation!

§ Approximate Systems Topologies § Approximate Computing Mechanisms – Approximate accelerators. – Approximate (Analog) Computing & other logic – Relaxed Cache Coherency & devices Consistency – Approximate memory cells (SRAM / DRAM / PCM) – Approximate Storage § Approximate Systems & Circuit – Stochastic Computing techniques Topologies – Neuromorphic Computing – Approximate Multipliers & Adders – Approximate Circuit Topologies § Stochastic Information Processing – Synthesis of Approximate Circuits Fabrics – CNT – RRAM / PCM etc

DAMMIT, AND IF YOU WANT TO“Nature MAKE isn’t classical, dammit, A SIMULATION OF NATURE, YOU’Dand if you want to make a simulation of nature, you’d BETTER MAKE IT QUANTUM better make it quantum MECHANICAL, AND BY GOLLY,mechanical, IT’S A and by golly, it’s a wonderful problem, because it WONDERFUL PROBLEM, BECAUSEdoesn’t IT look so easy.” DOESN’T LOOK SO EASY.” -Richard P. Feynman - RICHARD P. FEYNMAN

1. Quantum measurement: observing a state disturbs it, and returns only partial information about the state (uncertainty principle).

2. Quantum entanglement: Two systems can exist in an entangled state, causing them to behave in ways that cannot be explained by supposing that each particle has some state of its own.

aluminum ~1nm barrier,

Al2O3 aluminum

Opcal micrograph of a Josephson juncon is a non-linear inductor “transmon” superconducng circuit element qubit fabricated at IBM Circuit elements ‘anharmonic’ oscillator

resistor Josephson junction

capacitor inductor Use microwaves (5 Gigahertz) to control and operate © 2015 INTERNATIONAL BUSINESS MACHINES CORPORATION 11 I.B.M. Inch Closer on Quantum Computer - The New York Times http://www.nytimes.com/2012/02/28/technology/ibm-inch-close...

Coherence IBM Busts Record for 'Superconducting' Quantum Computer |... http://www.wired.com/2012/02/ibm-quantum-milestone/ COHERENCE PROGRESSION

TECHNOLOGY

10000.000 I.B.M. Researchers Inch Toward Quantum Computer Data Visualization By KENNETH CHANG FEB. 28, 2012 I.B.M. is jumping into an area of computing that has, until now, been primarilyand the Discovery province of academia: the quest to build a quantum computer. Qlik® delivers intuitive platform solutions 1000.000 for self-service data visualization, guided A computer that took advantage of the oddities of quantum physics couldanalytics solve applications, embedded in seconds certain problems that would occupy present-day computers for billionsanalytics of and reporting to approximately 36,000 customers worldwide. years. But for now, it is impossible to build such a computer because the bitsCompanies of of all sizes, across all information it would need for the calculations fall apart before a calculation canindustries be and geographies, use Qlik solutions to visualize and explore completed. The problem is, in essence, like trying to knit a sweater with yarn that 100.000 information, generate insight and make unravels before the first purl. better decisions. Here, check out how data exist in the digital world around us On Tuesday, I.B.M. researchers will present experimental results that theyevery say day. put them close to solving this problem, both by lengthening the lifetime of theVisit Qlik.com for more. quantum bits of information and by quickening the pace of computation. The 10.000 presentation will take place at a meeting of the American Physical Society in Boston.

“In the past, people have said, maybe it’s 50 years away, it’s a dream, maybe it’ll happen sometime,” said Mark B. Ketchen, manager of the physics of information CADE METZ BUSINESS 02.28.12 6:30 AM group at I.B.M.’s Thomas J. Watson Research Center in Yorktown Heights, N.Y. “I 1.000 used to think it was 50. Now I’m thinking like it’s 15 or a little more. It’s within reach. It’s within our lifetime. It’s going to happen.”

IBMMany BUSTS university researchers RECORD have done good work FOR solving the basic science Microseconds problems, Dr. Ketchen said, but “it’s going take an I.B.M. in the end to put it 0.100 ‘SUPERCONDUCTING’ QUANTUM 1 of 3COMPUTER 11/11/15, 11:32 AM 0.010

0.001 1 of 10 11/11/15, 12:01 AM 1999 2005 2007 2010 2012 2020 © 2015 INTERNATIONAL BUSINESS MACHINES CORPORATION 12 The IBM Quantum Experience Launched May 4, 2016 Users from 113 countries around the world

5 qubit processor on the cloud Composer

Fully fault-tolerant universal quantum computers are sll a long way away, but soon beyond 50 qubits, we will be in a post-quantum world

Quantum Quantum Condensed Maer Quantum enhanced Chemistry Chromodynamics Physics opmizaon

Molecule simulaon, Hadron simulaon, high bond properes, High Tc e.g. Travelling energy physics, nuclear electronic structure superconducng salesman-like physics materials problems (constrained sasfacon) Through IBM enabled by the Quantum Experience, become a part of this new age of discovery

Cryptography Designing Better Drugs Machine Learning Searching Big Data & New Materials

Thank You!