Intel 2020 Labs Day Keynote Slides
Download as PDF, PPTX0 likes89 views
The document discusses disruptive research at Intel Labs aimed at revolutionizing computing over the next decade, focusing on data generation and analysis challenges. Key topics include advancements in integrated photonics, neuromorphic computing, and quantum computing technologies, highlighting their potential to enhance performance and energy efficiency compared to traditional computing methods. Additionally, the document addresses privacy and collaboration in AI through federated learning and other emerging computing paradigms.
More Related Content
Similar to Intel 2020 Labs Day Keynote Slides (20)
Intel 2020 Labs Day Keynote Slides
- 2. Labs Day 2020 Disruptive Research for the Next Decade of Computing RichUhlig Intel Sr. Fellow, VP, & Directorof Intel Labs
- 4. labs
- 5. labs TheDataProblem We are generatingdataat a fasterrate than our abilityto analyze,understand,transmit, secure andreconstructin real-time Zettabytes 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 150 100 50
- 6. labs PerformanceDemocratization Compute 1980 1990 200 0 2010 2020 NetworkEverything 1018 109 104 1015 102 CloudEverything Mobile Everything Digitize Everything For Everyone Exascale 1 0 0 B I N T E L L I G E N T C O N N E C T E D D E V I C E S Distributed Intelligence
- 7. labs
- 9. labs Light Source Guide Light Photodetection Building Blocks Modulation 011010100100101000101110101001 010100010110111010010010 Low Cost Assembly
- 10. labs Other names and brands may be claimed as the property of others
- 12. labs
- 13. Labs Day 2020 JamesJaussi Senior Principal Engineer and Director of PHY Research Lab
- 14. labs Thousands >10km Millions 100m Billions 1m Trillions <1cm OPTICALELECTRICAL
- 15. labs Over Intel 100G transceivers shipped to our customers
- 16. labs Integrated Photonics Bringingoptical I/O directly into our servers and onto our packages
- 18. labs
- 20. labs Credit: DARPA for trendlines 1990 2000 2010 2020 2030 1 10 100 1000 10000 Total Power per Package Power for off-chipI/O Power(W) Trendlines POWER WALL Year
- 21. labs Integrated Photonics will fundamentally change compute I/O.
- 23. labslabs Silicon Photonics CMOS Electronics Package Integration Detection Light Source Modulation Mux/Demux CMOS Circuits Guide/Amplify Light IntegratedPhotonics Packaging
- 24. labs Industry-leading Prototype with Key Technology BuildingBlocks IntegratedPhotonicsPrototype labs 4 All-silicon Photodetectors Integrated Semiconductor OpticalAmplifier Electronic CMOSIC Integrated Laser 4 Micro-ring Modulators
- 25. labs
- 26. Labs Day 2020 MikeDavies DirectorOf NeuromorphicComputing Research
- 27. labs 2018IROSDroneRacingCompetition Other names and brands may be claimed as the property of others
- 28. labs BrainsareUnrivaledComputingDevices Navigates and learns unknownenvironments at 22 mph Brain Power: 50 mW Mass: 2.2 grams Can learn to speak English words Can learn to manipulatecups fordrinking Pre-trained to fly between known gates at 5.6 mph Can’t learn anything online CPU/GPUcontroller Power: 18,000 mW Mass: ~40 grams Sources:PNAS, June 13, 2016; https://link.springer.com/article/10.1007/s00360-011-0603-1;Davide Scaramuzza, ETH Zurich andA. Loquercio et al, “DeepDrone Racing:From Simulation to Reality with Domain Randomization,” IEEETrans. Robotics, 2020.
- 29. labs A NewKindofComputerArchitecture StandardComputing ParallelComputing NeuromorphicComputing PARALLEL DENSECOMPUTE OFFLINETRAININGUSING LABELEDDATASETS PROGRAMMINGBY ENCODINGALGORITHMS SEQUENTIALTHREADS OFCONTROL SYNCHRONOUS CLOCKING SYNCHRONOUS CLOCKING PARALLEL SPARSECOMPUTE LEARNON THEFLY THROUGH NEURONFIRINGRULES ASYNCHRONOUS EVENT-BASEDSPIKES
- 30. labs Loihi KEYPROPERTIES ▪ 128k neurons and 128million synapses ▪ Compute-memory integrated architecture ▪ Fully digital in standard 14nm process ▪ Asynchronous design enables scalability ▪ Versatile on-chip learning – a firstfor the field Yet, ▪ No floating point numbers! ▪ No multiply-accumulators! Fundamental to deep learning hardware
- 31. labs Loihi KEYPROPERTIES ▪ 128k neurons and 128million synapses ▪ Compute-memory integrated architecture ▪ Fully digital in standard 14nm process ▪ Asynchronous design enables scalability ▪ Versatile on-chip learning – a firstfor the field Yet, ▪ No floating point numbers! ▪ No multiply-accumulators! Fundamental to deep learning hardware
- 32. labs Efficiency Precision Speed Capacity per dollar Neuromorphic Conventional (unbatched) LoihiEfficiency
- 34. labs
- 35. labs Other names and brands may be claimed as the property of others
- 36. labs INRCincludes over100groups Other names and brands may be claimed as the property of others
- 37. labs OpportunityatAllScales Visual Intelligence Personalized Computing (Real-timespeech,speaker ID, localization,denoising) At-ScaleProblem Solving Data Analytics,Security,Scientific Computing Robotic Sensing + Control SWaP-constrained AI (Autonomous systems) Intelligent Sensors (Lowlatency,event-based, anomalydetection) Human-ComputerInterfacing (EEG,neuroprosthetics) IoT Other names and brands may be claimed as the property of others
- 38. labs Other names and brands may be claimed as the property of others
- 39. labs Other names and brands may be claimed as the property of others
- 40. labs Constrained optimization problems (e.g.Sudoku) 1000x lower energy 100x faster Head direction localizationand learning v00x lower power vs CPU 1000x lower energy 100x faster Similarity search 24x faster and 30x lower energy(vs CPU) Visual-tactile sensing 45x lower power 20% faster (vs GPU) Adaptiverobotic armcontrol 40x lower power , 50% faster (vs GPU)
- 41. labs Other names and brands may be claimed as the property of others
- 42. labs
- 43. Labs Day 2020 AnneMatsuura Directorof Quantum Computing Architecture
- 44. labs Changingthe World Drug Design Financial Modeling “Quantum Will Change Everything” Climate Modeling Travel& Logistics Cryptography Other names and brands may be claimed as the property of others
- 45. labs QuantumComputing:KeyConcepts Superposition Classical Physics Quantum Physics Heads OR Tails Heads ANDTails Entanglement N Quantum Bits or Qubits= 2N States 1) 2) 3) 4) Observation or noise causes loss of information Fragility
- 50. labs A QuantumBit Google, IBM, Rigetti, DWave Honeywell, IonQ Intel Corporation, HRL Only one of these Qubits is built on the technologyof transistors Other names and brands may be claimed as the property of others
- 51. labs Other names and brands may be claimed as the property of others
- 52. labs Architecture:CompletelyNewKindofCompute ControlElectronics Quantum Algorithms QubitChip QuantumRuntime QuantumCompiler QubitControlProcessor Keysystemchallengesfor QuantumPracticality • New execution model • Error mitigation & resilience • Scalability • Interconnectcomplexity • Qubit device design
- 53. labs Puttingit AllTogether * *Q-NEXT brings togethernearly 100 world-class researchers from three national laboratories, 10universities and 10 leading U.S. technologycompanies withthe single goal of developing thescienceand technologyto control and distribute quantum information. ControlElectronics QuantumAlgorithms QubitChip QuantumRuntime QuantumCompiler QubitControlProcessor Full-StackResearchTestbed
- 54. labs
- 55. labs The Data Silo Problem ▪ Privacy / Legality ▪ Data too valuable ▪ Data too largeto transmit
- 56. Labs Day 2020 JasonMartin PrincipalEngineer,Securityand PrivacyResearch
- 57. labs
- 58. labs
- 59. labs Aggregation Server Training Infrastructure Hospital A Hospital B Hospital C Hospital A FederatedLearning– MoveComputetoData Centralized Learning Model Data A Data B Data C Updated model Model Update A Model Update B Model Update C Federated Learning Hospital B Hospital C
- 60. labs Sheller, M.J., Edwards, B., Reina, G.A. etal. Federatedlearningin medicine:facilitatingmulti-institutional collaborations without sharingpatient data. SciRep10, 12598(2020). Intel-UPennCollaboration Howmuch better does each institution do when training on the full data vs. just their own data? 17% BETTER 2.6% BETTER on their own validation data on the hold-outBraTS data Other names and brands may be claimed as the property of others
- 64. labs
- 65. labs FHEPerformanceExplorations PlaintextPerformance 101 102 103 104 105 106 1 Exploration See cited sources forworkloads and configurations. Results mayvary. Jung et al. HEANN Demystified. arXiv:2003.04510, March2020 Riazi et al. HEAX. ASPLOS 2020
- 66. labs Intel Launches Private AI Collaborative Research Institute Advancingtechnologiesinprivacy andtrustfordecentralizedAI Carnegie Mellon University National University of Singapore Université Catholique de Louvain University of California, San Diego University of Louvain University of Southern California University of Toronto University of Waterloo Technische Universität Darmstadt Other names and brands may be claimed as the property of others
- 67. labs
- 68. labs ProgrammingChallenges 10,000 1,000 100 10 Original Performance A cosmology application from the Stephen Hawking Institute Runtime(seconds) Source:Intel Labs Ninja – an expert in SW development generallyrequiringadeepunderstandingof HW Programmingin the XPUEra
- 69. labs ProgrammingChallenges 10 monthsof work/training froma dedicated Ninja workingon-site Optimization over original 10,000 1,000 100 10 Original Performance Performance from Ninja A cosmology application from the Stephen Hawking Institute Runtime(seconds) Source:Intel Labs Ninja – an expert in SW development generallyrequiringadeepunderstandingof HW Programmingin the XPUEra
- 70. labs Million Million 1990 202020102000 ProgrammerPopulation C Ada Pascal C Java JavaScript Java JavaScript PHP JavaScript Python Java Top ProgrammingLanguages 56% 34% 24% 32% 25% 20% 32% 20% 17% 24% 23% 22% Increase in 6 years (2013-2019) ProgrammingChallenges Other names and brands may be claimed as the property of others
- 71. Labs Day 2020 JustinGottschlich PrincipalEngineer,Machine Programming Research
- 72. labs Howdoyoudevelopsoftwaretofullyexercisethe capabilitiesof thesenoveltypesofhardware? • Any technique that automates software development • Democratize and accelerate the creation of quality software
- 73. labs The Three Pillars of Machine Programming
- 74. labs Less than 1% of the global population can code.
- 75. labs Empower all people to express their intentionto a machine and have the machine build the software
- 76. labs Programmers spend 50% of their time debugging Can we do better?
- 77. labs Legal Information Performancevaries by use, configuration and otherfactors. Learnmoreat www.Intel.com/PerformanceIndex. Performanceresults are based on testingas of dates shown in configurations and maynot reflect all publicly available updates. See backup for configuration details. No product or component canbe absolutely secure. Yourcosts and results mayvary. Resultshave beenestimatedorsimulated. Inteltechnologies mayrequireenabled hardware,software or service activation. Inteldoes not controlor audit third-partydata. Youshould consult othersources toevaluateaccuracy. Inteldisclaims all express and implied warranties, including without limitation,theimplied warranties of merchantability, fitness for a particular purpose, and non- infringement,as well as any warrantyarising from course of performance, course of dealing, orusage in trade. © Intel Corporation. Intel,theIntellogo, and otherIntelmarks aretrademarks of IntelCorporationor its subsidiaries. Othernamesand brands may be claimed as the propertyof others.