No Slide Title

Download Report

Transcript No Slide Title

System 2020:
Research Grand Challenges in
Computer Architecture
Mary Jane Irwin
Penn State University
John Shen
Intel
1
What is the next big thing ?
Mainframes
Mini’s
Workstations
PC’s
Eniac
???
2
What are the mega trends ?
1. Wired  Wireless
 Telecommunication
 Internet/Computing
2. Patch-work Wireless  Blanket Wireless
3. Personal Computer  Mobile Computer 
Persistent/Transparent Computer
4. Embedded vs. High-end  Convergence?
5. Sever vs. Client  Convergence?
3
And anticipated usage models ?
Human-centric:

Intelligent spaces

Personal agents
For work, education, leisure, entertainment
Active displays, sensory devices, immersive experience
Feature rich gadgets; useful real-time information
Highly mobile, roam seamlessly from space to space
Infrastructure-centric:

Traditional server farms and data centers

Fabric for supporting human-centric uses
Very large scale information fusion, storage, analysis
Communication and synchronization between spaces
Proactively pushing information to roaming agents
Support enormous number of distributed and roaming “servers”
4
The computing paradigm ala Google
5
The computing paradigm ala Nokia
6
What are the components of a GC?
 A “grand” scale problem that will require
at least a decade of concentrated
research to make substantive progress
1. that has a measurable outcomes/milestones,
2. that will excite and engage the computer
architecture research community,
3. and that is deserving of considerable
investment by funders because it will
materially advance the capabilities and
conduct of society.
7
1W Featherweight Supercomputer
1. For the goal of 1TOP/W will need 250 to 1000X
improvement in performance/W
 1TOP/W = .001 nJ/op vs today’s ~30nJ/op
2. Architects are already engaged
3. Funding and impacts
 societal impacts are clear and compelling:
pervasive intelligent sensors, embedded
supercomputing appliances, . . .
 funding investments?
8
Featherweight Challenges
 power/energy reductions
dynamic and leakage, HW/SW mode controls, . . .
 technology issues (65nm45nm32nm)
↑ process variation, ↑ transient/aging faults,
advanced packaging (SoC  MCP  3D), . . .
 design issues
cost, design time & tools, verification & test, . . .
 performance improvements
CMPs & SMT, heterogeneous cores,
programmable accelerators, eDRAMs, NoCs, . . .
 programmability . . .
9
Popular Parallel Programming (P3)
1. Software and architecture support that makes
parallel programming easy
 If 2X per 2 year perf. gains continue, will soon have
1000-way chip-level parallelism
2. Architects are becoming engaged but can’t do the
job alone
 need compiler, system & application developers
3. Funding and impacts
 a necessary enabling technology for future chips
(e.g., the 1W Featherweight Supercomputer)
 funding investments?
10
P3 Challenges
 new programming languages/models
that are correct, efficient, scalable, portable, . . .
that require minimal exposure of the programmer
to low-level details
and that support multi-modal parallelism
data-parallel, embarrassingly parallel, irregularly parallel
 microarchitecture support
lightweight thread/process communication and
synchronization, monitoring for reliability and
thermal hot spots, dynamic adaptation, . . .
 development support
benchmarks, prototyping platforms, tools for
debugging, performance tuning, . . .
11
Dependable Systems
1. Self-healing, trustworthy hardware and software
systems everywhere
 Low-cost computing you can trust your life on
 2x improvement in mean work-to-failure per generation
 Cost of ownership, vendor costs for liability/repair
2. Architects already engaged but can’t do the job
alone
 A system stack problem – devices, circuits, languages,
OS, applications, dependability analysts
3. Funding and impacts
 The s/w problem alone is ~ 0.6% GDP of the US
 funding investments?
12
Dependable Systems Challenges
Host of hardware reliability problems
Transient, aging, infant mortality, variations,…
Software reliability, security getting worse
Steep constraints
Area, power, perf (even for high-end systems)
Architects can provide low cost solutions
Workload-aware, selective, fast, adaptive
Bring dependability to h/w-s/w interface
Adapt to trade off reliability, security, perf, power
Integrated cross-layer solution from devices to app
13
New Computing Models
1. Beyond the stored program architecture
 data flow? neural network?
2. “Expanding the box” for architects
 neuroscientists, biologists, chemists, . . .
3. Funding and impacts
 neuro-prosthetics, telepathy, . . .
 funding investments?
14
“Brain” Challenges
High risk – but high payoff
Neuroscientists are a long way from unraveling the
mysteries of the neocortex
Take partial steps – augment certain brain functions
(hearing for the deaf, vision for the blind, mobility
for the quadrapeligic),
Take advantage of emerging technologies
Heterogeneous systems: silicon + nanosensors and
actuators, emerging nanotechnologies (CNT, QCAs,
quantum, . . .)
15
Watch for the final report
http://www.cra.org/Activities/grand.challenges
/architecture/home.html
And check out the reports from the previous
Grand Challenges conferences
http://www.cra.org/grand.challenges/
16