21st Century Computer Architecture A community white paper 21st Century Computer Architecture A...
If you can't read please download the document
21st Century Computer Architecture A community white paper 21st Century Computer Architecture A community white paper NSF Outbrief on June 22, 2012 Information
21st Century Computer Architecture A community white paper 21st
Century Computer Architecture A community white paper NSF Outbrief
on June 22, 2012 Information & Commun. Techs Impact
Semiconductor Technologys Challenges Computer Architectures Future
Pre-Competitive Research Justified Process, Participants &
Backups
Slide 2
20 th Century ICT Set Up Information & Communication
Technology (ICT) Has Changed Our World o Required innovations in
algorithms, applications, programming languages, , & system
software Key (invisible) enablers (cost-)performance gains o
Semiconductor technology (Moores Law) o Computer architecture (~80x
per Danowitz et al.) 2
21 st Century Promise ICT Promises Much More o Data-centric
personalized health care o Computation-driven scientific discovery
o Human network analysis o Much more: known & unknown
Characterized by o Big Data o Always Online o Secure/Private oo
Whither enablers of future (cost-)performance gains? 4
Slide 5
Technologys Challenges 1/2 Late 20 th CenturyThe New Reality
Moores Law 2 transistors/chip Transistor count still 2 BUT Dennard
Scaling ~constant power/chip Gone. Cant repeatedly double
power/chip 5
Slide 6
Classic CMOS Dennard Scaling: the Science behind Moores Law 6
National Research Council (NRC) Computer Science and
Telecommunications Board (CSTB.org) Scaling: Oxide: t OX / Results:
Power Density: Voltage: V/ Power/ckt: 1/ 2 ~Constant (Finding 2)
Source: Future of Computing Performance: Game Over or Next Level?,
National Academy Press, 2011
Slide 7
Power Density: ~Constant Post-classic CMOS Dennard Scaling 7
National Research Council (NRC) Computer Science and
Telecommunications Board (CSTB.org) Scaling: Oxide: t OX / Results:
Voltage: V/ V Power/ckt: 1 22 1/ 2 Post Dennard CMOS Scaling Rule
TODO: Chips w/ higher power (no), smaller ( ), dark silicon ( ), or
other (?)
Slide 8
Technologys Challenges 2/2 Late 20 th CenturyThe New Reality
Moores Law 2 transistors/chip Transistor count still 2 BUT Dennard
Scaling ~constant power/chip Gone. Cant repeatedly double
power/chip Modest (hidden) transistor unreliability Increasing
transistor unreliability cant be hidden Focus on computation over
communication Communication (energy) more expensive than
computation 1-time costs amortized via mass market One-time cost
much worse & want specialized platforms 8 How should architects
step up as technology falters?
Slide 9
21 st Century Comp Architecture 20 th Century21 st Century
Single-chip in generic computer Architecture as Infrastructure:
Spanning sensors to clouds Performance plus security, privacy,
availability, programmability, Cross- Cutting: Break current layers
with new interfaces Performance via invisible instr.-level
parallelism Energy First Parallelism Specialization Cross-layer
design Predictable technologies: CMOS, DRAM, & disks New
technologies (non-volatile memory, near-threshold, 3D, photonics, )
Rethink: memory & storage, reliability, communication 9 X
X
Slide 10
21 st Century Comp Architecture 20 th Century21 st Century
Single-chip in stand-alone computer Architecture as Infrastructure:
Spanning sensors to clouds Performance plus security, privacy,
availability, programmability, Cross- Cutting: Break current layers
with new interfaces Performance via invisible instr.-level
parallelism Energy First Parallelism Specialization Cross-layer
design Predictable technologies: CMOS, DRAM, & disks New
technologies (non-volatile memory, near-threshold, 3D, photonics, )
Rethink: memory & storage, reliability, communication 10
Slide 11
What Research Exactly? Example o Dream of sensor/phone apps
that burn >> 1W package o Note: Short-term use often followed
by long idle period o Exploit Computational Sprinting (to idle!) 16
cores (32W) for 100s ms in 1W package Use phase-change material
thermal capacitor Research areas in white paper (& backup
slides) 1.Architecture as Infrastructure: Spanning Sensors to
Clouds 2.Energy First 3.Technology Impacts on Architecture
4.Cross-Cutting Issues & Interfaces Computation Sprinting
funded NSF CCF-1161505 Much more research developed by future PIs!
11
Pre-Competitive Research Justified Retain (cost-)performance
enabler to ICT revolution
http://cra.org/ccc/docs/init/21stcenturyarchitecturewhitepaper.pdf
Successful companies cannot do this by themselves o Lack needed
long-term focus o Dont want to pay for what benefits all o Resist
transcending interfaces that define their products 13
Slide 14
White Paper Process Late March 2012 o CCC contacts coordinator
& forms group April 2012 o Brainstorm (meetings/online doc) o
Read related docs (PCAST, NRC Game Over, ACAR1/2, ) o Use online
doc for intro & outline then parallel sections o Rotated
authors to revise sections May 2012 o Brainstorm list of researcher
in/out of comp. architecture o Solicit researcher
feedback/endorsement o Do distributed revision & redo of intro
o Release May 25 to CCC & via email Kudos to participants on
executing on a tight timetable 14
Slide 15
White Paper Participants * contributed prose; ** effort
coordinator Thanks of CCC, Erwin Gianchandani & Ed Lazowska for
guidance and Jim Larus & Jeannette Wing for feedback 15 Sarita
Adve, U Illinois * David H. Albonesi, Cornell U David Brooks,
Harvard U Luis Ceze, U Washington * Sandhya Dwarkadas, U Rochester
Joel Emer, Intel/MIT Babak Falsafi, EPFL Antonio Gonzalez,
Intel/UPC Mark D. Hill, U Wisconsin *,** Mary Jane Irwin, Penn
State U * David Kaeli, Northeastern U * Stephen W. Keckler,
NVIDIA/U Texas Christos Kozyrakis, Stanford U Alvin Lebeck, Duke U
Milo Martin, U Pennsylvania Jos F. Martnez, Cornell U Margaret
Martonosi, Princeton U * Kunle Olukotun, Stanford U Mark Oskin, U
Washington Li-Shiuan Peh, M.I.T. Milos Prvulovic, Georgia Tech
Steven K. Reinhardt, AMD Michael Schulte, AMD/U Wisconsin Simha
Sethumadhavan, Columbia U Guri Sohi, U Wisconsin Daniel Sorin, Duke
U Josep Torrellas, U Illinois * Thomas F. Wenisch, U Michigan *
David Wood, U Wisconsin * Katherine Yelick, UC Berkeley/LBNL *
Slide 16
Back Up Slides Detailed research areas in white paper
1.Architecture as Infrastructure: Spanning Sensors to Clouds
2.Energy First 3.Technology Impacts on Architecture 4.Cross-Cutting
Issues & Interfaces
http://cra.org/ccc/docs/init/21stcenturyarchitecturewhitepaper.pdf
Findings on National Academy Game Over Study Glimpse at DARPA/ISAT
Workshop Advancing Computer Systems without Technology Progress
16
Slide 17
1. Architecture as Infrastructure: Spanning Sensors to Clouds
Beyond a chip in a generic computer To pillar of 21 st century
societal infrastructure. o Computation in context (sensor, mobile,
, data center) o Systems often large & distributed o
Communication issues can dominate computation o Goals beyond
performance (battery life, form factor) Opportunities (not
exhaustive) o Reliable sensors harvesting (intermittent) energy o
Smart phones to Star Treks medical tricorder o Cloud infrastructure
suitable for both Big Data streams & low-latency
qualify-of-service with stragglers o Analysis & design tools
that scale 17
Slide 18
2. Energy First Beyond single-core performance computer To
(cost-)performance per watt/joule Energy across the layers o
Circuit/technology (near-threshold CMOS, 3D stacking) o
Architecture (reducing unnecessary data movement) o Software
(communication-reducing algorithms) Parallelism to save energy o
Vast (fined-grained) homogeneous & heterogeneous o Improved SW
stack o Applications focus (beyond graphic processing units)
Specialization for performance & energy efficiency o
Abstractions for specialization (reducing 1-time cost) o
Energy-efficient memory hierarchies o Reconfigurable logic
structures 18
Slide 19
3. Technology Impacts on Architecture Beyond CMOS, Dram, &
Disks of last 3+ decades to Using replacement circuit technologies
o Sub/near-threshold CMOS, QWFETs, TFETs, and QCAs Non-volatile
storage o Beyond flash memory to STT-RAM, PCRAM, & memristor 3D
die stacking & interposers o logic, cache, small main memory
Photonic interconnects o Inter- & even intra-chip Design
automation o from circuit-design w/ new technologies to o pre-RTL
functional, performance, power, area modeling of heterogeneous
chips & systems 19
Slide 20
4. Cross-Cutting Issues & Interfaces Beyond performance w/
stable interfaces to New design goals (for pillar of societal
infrastructure) o Verifiability (bugs kill) o Reliability
(dependability computing base?) o Security/Privacy (w/ non-volatile
memory?) o Programmability (time to correct-performant solution)
Better Interfaces o High-level information (quality of service,
provenance) o Parallelism ((in)dependence, (lack of) side-effects)
o Orchestrating communication ((recursive) locality) o
Security/Reliability (fine-grain protection) 20
Slide 21
Executive summary (Added to National Academy Slides) Highlights
of National Academy Findings (F1) Computer hardware has
transitioned to multicore (F2) Dennard scaling of CMOS has broken
down (F3) Parallelism and locality must be exploited by software
(F4) Chip power will soon limit multicore scaling Eight
recommendations from algorithms to education We know all of this at
some level, BUT: Are we all acting on this knowledge or hoping for
business as usual? Thinking beyond next paper to where future value
will be created? Questions Asked but Not Answered Embedded in NA
Talk Briefly Close with Kbler-Ross Stages of Grief: Denial
Acceptance Source: Future of Computing Performance: Game Over or
Next Level?, National Academy Press, 2011 Mark Hill talk (
http://www.cs.wisc.edu/~markhill/NRCgameover_wisconsin_2011_05.pptx
)
Slide 22
The Graph 22 System Capability (log) 80s90s00s10s20s30s40s CMOS
Fallow Period New Technology Our Focus 50s Source: Advancing
Computer Systems without Technology Progress, ISAT Outbrief
(http://www.cs.wisc.edu/~markhill/papers/isat2012_ACSWTP.pdf) Mark
D. Hill and Christos Kozyrakis, DARPA/ISAT Workshop, March 26-27,
2012. Approved for Public Release, Distribution Unlimited The views
expressed are those of the author and do not reflect the official
policy or position of the Department of Defense or the U.S.
Government.
Slide 23
Surprise 1 of 2 Can Harvest in the Fallow Period! 2 decades of
Moores Law-like perf./energy gains Wring out inefficiencies used to
harvest Moores Law HW/SW Specialization/Co-design (3-100x) Reduce
SW Bloat (2-1000x) Approximate Computing (2-500x)
--------------------------------------------------- ~1000x = 2
decades of Moores Law! 23
Slide 24
Surprise 2 of 2 Systems must exploit LOCALITY-AWARE parallelism
Parallelism Necessary, but not Sufficient As communications energy
costs dominate Shouldnt be a surprise, but many are in denial Both
surprises hard, requiring vertical cut thru SW/HW 24