Transcript slides

“Single-chip Cloud Computer”
A many-core research platform from Intel Labs
Tor Lund-Larsen
Intel Labs
Agenda
 Introduction
 Motivation for the SCC
 Introducing the Intel “Single-Chip Cloud
Computer”
 Introducing the SCC Co-Traveler program
 Summary, Q & A
2
Intel in Braunschweig
 Intel acquisition in 2000
 Today: 100+: Research Scientist & Engineers,
Students
 Two main Intel groups on-site
– IAG: Intel Architecture Group - Product Development
– IL: Intel Labs – Research
3
One of the largest Intel R & D sites in Europe
Intel Labs Braunschweig
Research Charter:
• Emulation technology; accelerate IA architecture &
design
• “Tera-scale”/Many-Core processor prototyping
• Memory architectures; break the “memory wall”
• Joint research programs with European universities
Emulation infrastructure and
technology
4
Tera-scale
microprocessors
Memory architectures
Compute evolving to “TeraScale”
Entertainment, Learning
Performance
Financial Analytics
TIPS
ModelBased Apps
Personal Media Creation
and Management
GIPS
3D and
Video
MIPS
Health and
Medicine
Many-core
Multimedia
Multi-Core
KIPS
Text
Kilobytes
5
INTEL
TERA-SCALE
RESEARCH
Single-Core
Megabytes
Gigabytes
Dataset Size
Terabytes
Tera-Scale Scaling Challenges
Energy
Efficiency
6
Design
Complexity
Programming
Models
Emerging
Applications
Single-chip Cloud Computer
• Experimental many-core CPU/Platform for “Tera-Scale” HW/SW
research
• Many-core processor research  Hardware
• Parallel Programming research  Software
• Research platform shared with industry and academic collaborators to
enable/encourage “tera-scale” explorations
Energy
Efficiency
7
• Dynamic voltage/frequency scaling
• 1/3 power reduction for core-core I/O
Design
Complexity
• Array of small IA-based tiles could
lead to more agile, flexible designs
Programming
Model
• Message-passing approach proven
to scale to 1000’s processors
Application
Development
• Sharing with Microsoft* & others for
academic, industry innovation
The SCC Platform
• Debug tools – Memory Reader, SoftRAM, R/W SCC
Config Registers
• Provides SW based virtual I/O (e.g. performance widget)
• Konsole w/ SSH connections to all booted cores
• Comandline tools – sccBoot, sccReset, sccBMC,
sccKonsole
8
Intel SCC – a complete HW/SW “Tera-Scale” research
platform
A closer look
MEMORY
CONTROLLER
• 24 Dual-core tiles (48 IA cores)
• 24 Routers
• Mesh network with 256 GB/s
bisection bandwidth
• 4 Integrated DDR 3 memory
controllers
• 1.3 Billion transistors
9
R
R
1TIL
E
R
R
R
R
Dual-core SCC Tile
L2 Cache
ROUTER Message Buffer
L2 Cache
ROUTE
R
Core 1
Core 2
Dynamic Power Management
Fine-grain, software-controlled power management
8 Voltage and frequency islands  Dynamic range 25-125W
Each tile can run at a different frequency
6 banks of 4 tiles can run at different voltages
Independent V&F control for I/O network & MCs
Memory Controller
R
Tile
R
Memory Controller
R
Fn
V4R
R
Tile
R
Fn
Tile
R
Fn
Tile
V5R
Tile
R
Tile
R
Tile
Tile
R
V2R
Tile
Fn
Tile
Tile
R
Tile
R
R
Tile
10
V1R
Tile
Tile
R
Tile
Tile
R
Tile
R
Tile
R
V3R
V6R
Tile
R
Tile
48 IA cores at
25-125W
Tile
Tile
R
Memory Controller
Tile
Memory Controller




Advancing Parallel SW Research
 The SCC eliminates significant complexity & power by
removing hardware cache coherency
DATA
 Enables exploration of more scalable alternatives:
– Message passing models common in datacenter, HPC
– Software-managed, adaptive cache coherency
11
SCC Co-Traveler Program
Goal: Enable Tera-Scale Research by Industry/academic
institutions
 Access to SCC System(s), Tools, Documentation, Open Source SW, Support etc
 Access to “Eco-system” of users, Intel sponsored conferences, SCC Workshops
 Working with 100+ partners around the world
 Deployment in wavers, currently executing wave #1
"We're very excited about Intel's SCC. In the Barrelfish project we are designing
OS architectures for future multi-core and many-core systems. The chip's
memory system and message passing support are a great fit for us, and it's an
ideal vehicle for us to test and validate our ideas.“
– Prof. Timothy Roscoe, ETH Zürich
"The upcoming Single-chip Cloud Computer is of great interested to application
developers and tools researchers. The availability of the hardware will greatly
accelerate our development of applications and tools for massively parallel
computing platforms.”
– Prof. Wen-Mei Hwu, University of Illinois, UPCRC@Illinois co-director
12
SCC Co-Traveling in action
Financial Analytics
w/ shared virtual memory
JavaScript Physics Modeling
13
Microsoft Visual Studio
Advanced Power Management
HPC Parallel Workloads
Hadoop Web Search
SCC Co-Traveler Timeline
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Introduction Symposia/Workshops
Santa Clara
Feb 12
Germany
Mar 16
Workshops and Conferences
TBD
EWME
May 10-12
Research Proposal Process (1st wave)
Registration
Applications 15 Apr
Deadline
Notification
1st Week of May
Documentation/SW Availability
Overview, Messaging LIB,
EAS, How To Use Linux
Sample
Workflows
Final Docs
Website Active
Platform Availability
Beta
Testers
14
General HW Availability
Datacenter For Remote Access
Sep
Summary:
 Many-Core/Tera-Scale compute transition will happen!
 The Intel SCC is an experimental many-core research platform
designed to help addressing the “tera-scale” HW and SW challenges:
Energy
Efficiency
Design
Complexity
• Dynamic voltage/frequency scaling
• 1/3 power reduction for core-core I/O
• Array of small IA-based tiles could lead
to more agile, flexible designs
Programming • Message-passing approach proven to
scale to 1000’s processors
Model
Application
Development
• Sharing with Microsoft* & others for
academic, industry innovation
 Intel is making the SCC available world-wide to enable industry and
academic research and innovation.
15
You are invited to join us!
Thank You!
• Visit our website: www.intel.com/info/scc
•Technical Questions about SCC Platform, Research:
[email protected]
Copyright © 2010, Intel Corporation. All Rights Reserved.
Partner Disclaimer: It is acknowledged that the use of the word "Partner" is a commonly used term in the technology industry to designate a marketing
relationship between otherwise unaffiliated companies and is used in accordance with this common usage herein. The use of the word “Partner” herein
shall not be deemed to nor is it intended to create a partnership, agency, joint venture or other similar arrangement between Intel and such partners and
the employees, agents and representatives of one party shall not be deemed to be employees, agents or representatives of the other. Intel and the
partners shall be deemed to be independent contractors and shall have no authority to bind each other.
Intel and the Intel logo are trademarks of Intel Corporation in the United States and other countries.
* Other names and brands may be claimed as the property of others.
16
17
Acknowledgements
Jim Held, Jason Howard, Saurabh Dighe, Yatin Hoskote, Sriram
Vangal, David Finan, Gregory Ruhl, David Jenkins, Howard Wilson,
Nitin Borkar, Gerhard Schrom, Fabrice Pailet, Shailendra Jain, Tiju
Jacob, Satish Yada, Sraven Marella, Praveen Salihundam, Vasantha
Erraguntla, Michael Riepen, Guido Droege, Joerg Lindemann, Matthias
Gries, Thomas Apel, Kersten Henriss, Tor Lund-Larsen, Sebastian
Steibl, Shekhar Borkar, Vivek De, Rob Van Der Wijngaart, Timothy
Mattson
Extending Tera-scale Research
2006 Many-core Prototype
2009 Many-core Prototype
“Teraflops Research Processor”
“Single-chip Cloud Computer”
Many simple FP cores
Many fully-functional IA cores
Validated tiled-design concept
Prototypes a tiled-design microprocessor
Tested HW limits of a mesh network
Improved mesh with 3x performance/watt
Sleep capabilities at core and circuit level
Dynamic voltage & frequency scaling
Lightweight message passing
Message passing & controlled memory sharing
Limited programmability for basic benchmarks
Full programmability for application research
Primarily a circuit experiment
Circuit & software research vehicle
19
Intel Labs Braunschweig
 Design:
– IA core and the Message Passing solution
– DDR3 Memory Controllers design
 Validation:
– Logic validation of the complete chip
– FPGA Emulation for pre-silicon SW
prototyping
 Platform:
– Test bed system validation and bring up
platform
 Software:
– A Linux OS
– Platform firmware, operational SW & drivers
20