Paul Franzon North Carolina State University Raleigh NC paulfncsuedu 9195157351 httpwwwecencsueduerlfacultypaulfhtml Thanks Thanks to Rambus for hosting this event Spherically Gary Edge VP for Research and Jaimie Stuart for Logistics ID: 705765
Download Presentation The PPT/PDF document "IRDS Emerging Research Devices and Arch..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
IRDS Emerging Research Devices and Architectures NanoCrossbar Workshop
Paul FranzonNorth Carolina State UniversityRaleigh, NCpaulf@ncsu.edu919.515.7351http://www.ece.ncsu.edu/erl/faculty/paulf.htmlSlide2
ThanksThanks to Rambus for hosting this eventSpherically Gary Edge, VP for Research and Jaimie Stuart for LogisticsSlide3
BackgroundInternational Roadmap for Semiconductors (ITRS) used to sponsored by SIANow International Roadmap for Devices and Systems (IRDS) sponsored by IEEE
Emerging Research Devices (ERD) Chapter had a subsection called Emerging Research Architectures (ERA). BothEdited chapter in odd yearsHelp workshops in even yearsHeld a workshop on Storage Class Memory in 2012Been wanting to do workshops in other areas but lacked good definitionFor 2016 two workshopsNanocrossbarApproximate/ Stochastic/ Probabilistic computingSlide4
Core Group - ERAPaul Franzon, NCSU (editor)An Chen, IBM (ERA chair)Shamik
Das, MitreMatthew Marinella, SandiaErik DeBenidictis, SandiaGeoff Burr , IBMSlide5
Program-Centric
(performance and components dictated by designer)
Data-Centric
(performance and/or
components influenced
by
the data that is passed through the system
)
Good old-fashioned
Von NeumannNon-Von Neumann MemoryProcessorNon-VN Processor(including less-than-reliable VN)Trained off-lineTrained in-lineCMOSNon-CMOSCMOSNon-CMOSCMOSNon-CMOSCMOSNon-CMOSCMOSNon-CMOS
Deterministic/reliable
Non-
deterministic
SRAM
DRAM
Flash
NVM crossbarsfor S-SCM,M-SCM
Logic-in-memory
FPGA
NVM-basedFPGA
Coupledoscillators
True North
Execution of pre-trained ANN
OhmicWeave
SupervisedANN learning
CMOS
“Nextswitch”
New learning algorithms(unsupervised, reinforcement)
HTM
Probabilisticcomputing
Approximatecomputing
CMOSbeyond the designenvelope
Crossbarsfor STDP
Crossbarsfor backprop
GPUs
Coarse-GrainedReconfigurableArchitectures
Probabilistic Learning
RBM
Bayesian
TCAM
Analogcomputing
Analogcomputing(w/ Flash)
Quantumcomputing
Accelerators(multimedia, etc.)
NV computing
MLAcceler-ators(Convolution,SVM, ML)
AutomataProcessing
Active
InterconnectSlide6
2015 ChapterStarted tracking NanoCrossbars explicitlySlide7
Goals of WorkshopIdentify and quantify the state of the art in devices, design, modeling, fabrication, and employment of Nano-enabled Crossbars for computing.
Identify the research barriers impeding the use of NanoCrossbars for computing.Slide8
Questions asked to presentersGeneral, including memoriesWhat is the status of achieving linear repeatable response, low power, sufficiently long retention, fast writes, sufficiently distinguishable resistances in different states, and long write endurance in one nanoscale device?
Is the access device issue solved? What are the remaining issues?Slide9
… QuestionsNeuromorphic computingWhat are the requirements on device linearity, scalability and dynamic range?
What is achieved today?What are the tradeoffs exposed in achieving this?What style of non-traditional computing is best suited to nanodevice arrays? E.g. spiking neuron, deep network, full logic map, etc.Why?What are the specific gaps in device properties that are preventing us from achieving this paradigm?Slide10
… QuestionsAnalog computingWhat are the requirements on device linearity, scalability and dynamic range?
What is achieved today?What are the tradeoffs exposed in achieving this?What is the required device yield? What mechanisms are available for implementing working arrays in the presence of <100% yield?What levels of noise during readout can be tolerated?To what degree could closed-loop control (e.g., iterative resistance-tuning for higher accuracy) be available during device write?Slide11
Agenda0900– 0930 : Introduction: Paul Franzon, NC State University0930 – 1020 : Matt Marinella, Sandia
1020 – 1040 : Break1040 : 1120 : Geoff Burr IBM1120 – 1200 : Catchup1200 – 1300 : Lunch1340 : Dmitri Strukov, UCSB1420 : Kevin Cao, ASU 1420– 1440 : Break 1440 - 1520: Miao Hu, HPE
520
– 1600 : Wei Lu, UMich
1600
–
1640
:
Gert Cauwenberghs, UCSD (via webex)