upcrc overview - inteldownload.intel.com/.../upcrc/upcrcoverview_presentation.pdfjoint hw/sw r&d...
TRANSCRIPT
Andrew A. ChienVice President of Research
Intel Corporation
UPCRC Overview UPCRC Overview Universal Computing Research Centers
launched at UC Berkeley and UIUCUniversal Computing Research Centers
launched at UC Berkeley and UIUC
2
Announcement Key Messages
• Microsoft and Intel are announcing the establishment of two “Universal Parallel Computing Research Centers” with the first locations at the University of Illinois at Urbana-Champaign and the University of California at Berkeley.
• Under this alliance, Microsoft and Intel have committed to invest a combined $20 million in the center over the next five years
• The center will explore the next generation of hardware and software for parallel computing and enable a revolutionary change in the way people use technology.
3
Multi-core is now mainstream
0%
25%
50%
75%
100%
Q3’05 Q4’05 Q1’06 Q2’06 Q3’06 Q4’06
SINGLESINGLE--CORE*CORE*
MULTIMULTI--CORECORE
Market Crossover to multi-core in 2006
Market Crossover Market Crossover to multito multi--core in 2006 core in 2006
Over 6 MUOver 6 MUQuadQuad--core core
In 2007In 2007
Mul
ti-co
re T
op to
Bot
tom
Motivated by a need for better performance/watt
Source: IntelSource: Intel
4
More, better More, better transistors transistors
More coresMore cores
Continued Continued benefitsbenefitsfrom Moorefrom Moore’’s Laws Law
Moore’s Law Motivates Multi-CoreMoore’s Law Motivates Multi-Core
45nm45nm
++
20072007
105
103
107
109
Source: IntelSource: Intel
5
A Shift to Many CoresParallelism will enable tera-scale performance for new apps
TerabytesTerabytes
TIPSTIPS
GigabytesGigabytes
MIPSMIPS
MegabytesMegabytes
GIPSGIPS
Perf
orm
ance
Perf
orm
ance
Dataset SizeDataset SizeKilobytesKilobytes
KIPSKIPS
Multi-Media
3D &Video
Text
Models
Rich Visual ComputingRich Visual Computing
HealthHealth
TeraTera--scalescaleMany coresMany cores
MultiMulti--corecore
SingleSingle--corecore
Turing Data into Turing Data into UnderstandingUnderstanding
Printer X42Printer X42
SensingSensing& Perception& Perception
6
• P Programming Effort ≤ Sequential Programming Effort• P Programming that doesn’t increase programming complexity
– Productivity, Modularity, Interactions, Performance Tuning, etc.
• P Programming Approaches that have “forward scalability”– Application implementation gets faster (more parallel) on
succeeding generations of hardware platforms– Scale data sets, scale output quality – more parallelism without
reprogramming or retuning
• P Programming implementation techniques that deliver high and robust parallel performance– Enable programming at a high level– Manage workload, algorithm, and data irregularlity– Manage hardware differences and irregularity
• HW Architecture Innovations that Support Parallel Programs
Wanted: Breakthrough Innovations in ParallelismWanted: Breakthrough Innovations in Parallelism
7
Intel: Making Parallel Computing Pervasive
EnablingEnablingParallelParallel
ComputingComputing
Academic Research
UPCRCs
SoftwareProducts
Multi-coreEducation
TBB Open SourcedTBB Open SourcedSTMSTM--Enabled Compiler on Enabled Compiler on Whatif.intel.comWhatif.intel.comParallel Benchmarks at Parallel Benchmarks at PrincetonPrinceton’’s PARSEC sites PARSEC site
Joint HW/SW R&D Joint HW/SW R&D program to enable program to enable
Intel products Intel products 33--7+ in future7+ in future
Intel Tera-scaleResearch
Academic research Academic research seeking disruptiveseeking disruptiveinnovations 7innovations 7--10+10+years out years out
Community and Experimental Tools
Wide array of leadingWide array of leadingmultimulti--core SW core SW
development tools & development tools & info available today info available today
MultiMulti--core Education Programcore Education Program-- 400+ Universities 400+ Universities -- 25,000+ students 25,000+ students -- 2008 Goal: Double this2008 Goal: Double this
IntelIntel®® Academic CommunityAcademic CommunityThreading for MultiThreading for Multi--core SW core SW communitycommunityMultiMulti--core bookscore books
8
Today: Universal Parallel Computing Research Centers
• Catalyze breakthrough research enabling pervasiveuse of parallel computing
Parallel Programming
Languages, Compilers,Runtime, Tools
Parallel Applications
For desktop, gaming, and mobile systems
Parallel Architecture
Support new generation of programming models
and languages
Parallel Sys. S/W
Performance scaling, memory Utilization, & power consumption
9
UPCRC Partners in Research
Professor David PattersonProfessor David PattersonUCB UPCRC DirectorUCB UPCRC Director
Prof. Marc Prof. Marc SnirSnir
Prof. Prof. WenWen--Mei Mei HwuHwu
UIUC UPCRC UIUC UPCRC CoCo--Directors Directors
Intel & Microsoft provide funding and guidanceIntel & Microsoft provide funding and guidance
Universities direct groundbreaking research Universities direct groundbreaking research
10
UPCRC Funding Strategy
AcademiaH/W & S/WCompanies
Leadership$$
Guidance
Innovation
Trained ResearchersGovernment
State GovsState Govs
Matching
Fund
Major investment in mainstream parallel programming
Intel + MicrosoftIntel + Microsoft
$20 Million over 5 years$20 Million over 5 years
Matching University InvestmentsMatching University Investments
University of Illinois: $8 mil (committed)University of Illinois: $8 mil (committed)
UC Berkeley: $7 mil (grant application)UC Berkeley: $7 mil (grant application)
11
Future of Parallelism ComputingFuture of Parallelism Computing
• For Software and Hardware– Parallelism in all computing systems– Dramatic new capabilities enabled by performance and
performance / unit power– Established and varied software models for portable
parallelism– Architecture support for Parallel Software and other
capabilities
12
BACKGROUND
13
Single Core Performance is Stagnant Single Core Performance is Stagnant
Time
0.01
1.01
2.01
3.01
4.01
Jan-
85Ja
n -86
Jan-
87Ja
n-88
Jan-
89Ja
n-90
Jan-
91Ja
n -92
Jan-
93Ja
n-94
Jan -
95Ja
n-96
Jan-
97Ja
n-98
Jan-
99Ja
n-00
Jan-
01Ja
n -02
Jan-
03Ja
n -04
Jan-
05Ja
n-06
Jan -
07Ja
n-08
Jan-
09
Freq
uenc
y (G
Hz)
Frequency limited by leakage and power. Transistor counts continue to increase.
Multi-core performance: P ~ Area
Transistor Count (area)P
erf
orm
an
ce
Gap driving us to Multi-core
Single Stream PerformanceP ~ √Area
Source: IntelSource: Intel
14
Many-core Research Questions AboundMany-core Research Questions Abound
• Cores– How many? What size?– Homogenous, Heterogeneous– Programmable, Configurable,
Fixed-function
• Chip-level– Interconnect: Topology,
Bandwidth– Coordination– Management
• Memory Hierarchy– # of levels, sharing, inclusion– Bandwidth, novel Technology– Integration/Packaging
• I/O Bandwidth– Silicon-based photonics– Terabit links
Source: CTWatchQuarterly, Feb 2007
Manycore Chips (circa. 2012)?
15
VirtualEnvironments
FinancialModeling
Media Search& Manipulation
Web MiningBots’
EducationalSimulation
ModelModel--Based ApplicationsBased Applications
Intel’s Tera-scale Research Vision
High BandwidthI/O & CommunicationsHigh BandwidthI/O & Communications
Stacked, Shared Memory
Stacked, Shared Memory
Scalable Multi-coreArchitectures
Scalable Multi-coreArchitectures
ThreadThread--AwareAwareExecution EnvironmentExecution Environment
Parallel ProgrammingParallel ProgrammingTools & TechniquesTools & Techniques
16
Copyright © Intel Corporation. *Other names and brands may be claimed as The property of others. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise or used for any public Purpose under any circumstance, without the express written consent and permission of Intel Corporation.