V - ECMWF

Transcription

V - ECMWF

ECMWF WS#12
NEC HPC Strategy and
Products
31 October 2006
Toshiyuki Furui
NEC Corporation
Contents
NEC’s HPC Activity
NEC’s HPC Strategy and Products
Toward the Future of HPC
Ｐage 2
© NEC Corporation 2006
NEC
Confidential
©
NEC Corporation
2006
2
Contents
Ｐage 3
NEC
Confidential
©
NEC Corporation
2006
3
Supercomputer Performance Gains
NEC Earth Simulator
FLOPS (Peak)
15
Peta 10
2002 40TF
Blue Gene/L
SX-8/8R
14
10
+75%/year
Earth
Simulator
Copyright JAMSTEC/Earth Simulator Center
NEC SX-2
13
10
SX-6
SX-5
1983 1.3GF
12
Tera 10
SX-4
11
NEC SX-8/SX-8R
Cray-1
10
1976
SX-3
10
10
SX-2
9
Giga 10
8
ENIAC
+47%/year
1946, 300FLOPS
10
CRAY-1
7
10
CDC6600
6
Mega 10
1970
Ｐage 4
'80
'90
NEC
Confidential
©
NEC Corporation
2006
2000
'10
4
No.1 Vector Supercomputer in the World
Over 1000 SX systems (cumulative orders)
Europe
(HPCE)
Canada
North America
Asia
iOuranos Consortium
iUniversity of Victoria
iUniversity of Toronto
Japan
South America
(NEC do Brasil)
Oceania
(NEC Australia)
Sales (units) by region
Example of Users
Europe
Japan
iThe Met Office (United Kingdom)
iMeteo France (France)
iDanish Meteorological Institute (Denmark)
iGerman Climate Computing Center (Germany)
iCHMI (Czech)
iUniversity of Stuttgart/HLRS (Germany)
iCNRS/IDRIS (France)
iSwiss Center for Scientific Computing (Switzerland)
iAerospace Laboratories
(Netherlands, Germany, France, Italy)
iTohoku University Information Synergy Center
iOsaka University Cyber Media Center
iNational Institute for Environmental Studies
iNational Institute for Fusion Science
iMeteorological Research Institute - JMA
iCentral Research Institute of
Electric Power Industry
iJapan Aerospace Exploration Agency
iToyota Central R&D Labs,. Inc.
iNissan Motors
Ｐage 5
NEC
Confidential
©
NEC Corporation
2006
Japan
50%
Europe Others
36% 14%
Asia Pacific/South America
iBureau of Meteorology / CSIRO (Australia)
iKorea Institute of Science, Technology and
Information (Korea)
iMeteorological Services of Singapore (Singapore)
iNational Institute for Space Research (Brazil)
5
SX Series in Weather/Climate Community
･ Victoria University
･ University of Toronto
･Danish Meteorological Institute (DMI)
･ Canadian Climate Research
･Deutsches Klimarechenzentrum (DKRZ)
Consortium
･Institute for Atmospheric Physics in Germany (IAP)
・ Met Office (UKMO)
･Czech Hydrometeorological Institute (CHMI)
・Interdisciplinary Center for Mathematical and
Computational Modeling, Warsaw University (ICM)
Europe
Japan
Meteo Swiss (CSCS)
Central Institute of Meteorology
and Geodynamics (ZAMG)
・Instituto Nazionale di Geofisica
e Vulcanologia (INGV)
Asia
Ѓњ
North America
･Earth Simulator
･National Institute of
Environmental Studies (NIES)
･Japan Marine Science and Technology
Center (JAMSTEC)
・Meteorological Research Institute (MRI)
・ Meteo France
Instituto
Nacional de
Pesquisas
Da Amazonia (INPA)
・Meteorological Service
Singapore (MSS)
Australia
・South African Weather
Service (SAWS)
Ｐage 6
･Bureau of Meteorology (BOM)/CSIRO
NEC
Confidential
©
NEC Corporation
2006
･Instituto Nacional De
Pesquisas Espaciais (INPE)
South America
6
Contents
Ｐage 7
NEC
Confidential
©
NEC Corporation
2006
7
NEC’s Computer Product Strategy
Leveraging two key technologies:
High performance technology and Highly reliability technology
High Reliability
Technology
High Performance
Technology
◎Most advanced technology
CPU Module
• High-speed/high-density VLSI
• High-density packaging
• High-efficiency cooling
• High-speed interconnect
◎Parallel processing technology
◎Cluster control technology
◎VALUMO
(Platform Technology)
• Autonomy/Virtualization
• Fault-tolerance
• Continuous operation
Technology
leader products
iExpress
SuperACOS
IPF Server/ Express
ft/Blade Network
Server
computer Mainframe Blade Server PC Server IA
Server
Ｐage 8
NEC
Confidential
©
NEC Corporation
2006
iStorage
DVD
8
NEC’s Strategy on HPC
SX: Vector supercomputer based HPC system
Overwhelming sustained performance and high reliability
with NEC original cutting edge technologies
Ever-increasing per Core(CPU) performance
Seamless connection with other servers
Right platform for right application,
which best fits the customers’ needs
SX
SX+Scalar
Scalar
IPF scalar server
PC cluster
Integrated NEC HPC solution
Ｐage 9
NEC
Confidential
©
NEC Corporation
2006
9
NEC’s HPC Products Lineup
Vector
Supercomputers
SX
Compact vector
server SXSX-8i
Supercomputer
SX Series
Other Platforms
GFS
Scalar Servers
(Global File System)
IPF
Scalar Server
TX7 Series
Express5800/
1020Ba
(IPF Blade)
IPF Blade Server
Rackmount
PC Cluster
PC Cluster
FC-SW
iStorage
EMC
Hitachi
Blade
(IA Server Cluster/IA Blade Server)
IA Workstation
Ｐage 10
Express5800/50
Series
NEC
Confidential
©
NEC Corporation
2006
10
SX Innovation
performance
Continual Break-Through Achieved
in both Technology and Architecture
Super large cluster
(>500 nodes)
Large cluster
(>100 nodes)
Architecture
SX-8R
SX-8
SX-8 SX-8R
SX-6
SX-6
Multi nodes
(>10nodes)
SX-5
SX-5
1 nodemodule
SX-4
SX-4
Multi Processor
Single node
>1GFLOPS
Single chip
vector processor
SX-3
SX-3
CMOS
Air-cooling
SX-2
SX-2
2cm
2cm
45.7cm
38.6cm
Bipolar
Water Cooling
1985
Ｐage 11
1990
Technology
CPU
1995
NEC
Confidential
©
NEC Corporation
2006
2000
2005
11
SX-8R Product Highlights
1. World’s fastest vector supercomputer
with maximum performance of 144 TFLOPS
- Very large scale : Up to 512 nodes, 4,096 CPUs
- Very large memory / memory bandwidth: 256TB / 288TB/s
- High speed data transfer between nodes : 8TB/s in total
2. High-density packaging with state-of-the-art technology
- Single-chip vector processor with 35.2GFLOPS performance
- Leading-edge CMOS technology with 90-nanometer process /copper
interconnects
- Single-module node with 281.6GFLOPS performance
3. Enhanced SUPER-UX / Tuned applications
- Proven operating system for SX series enhanced to expand scalability
- A lot of ISV application programs tuned for SX series available
Ｐage 12
NEC
Confidential
©
NEC Corporation
2006
12
SX-8R Enhancement
Vector adder and multiplier doubled
SX-8 : (Multiply + Add) x 4pipes x 2GHz = 16GF
SX-8R: (Multiply + Add) x 2sets x 4pipes x 2.2GHz = 35.2GF
Memory capacity is doubled
SX-8
SX-8R
128GB
64GB
256GB/node (DDR2 RAM)
128GB/node (FCRAM)
Clock-up
2.2GHz (10% up)
note: ONLY FOR DDR2 models
FCRAM model’s clock cycle remains with 2GHz.
Ｐage 13
NEC
Confidential
©
NEC Corporation
2006
13
Comparison of CPUs (SX-8, SX-8R)
SX-8
4 Vector
pipelines
Mask Reg.
SX-8R
4 Vector
pipelines
Mask
Mask Reg.
logical
Mult./Log
M/A V Pipe doubled
Multiply
Load
or
Store
Vector
Reg.
Load
or
Store
Add
Vector
Reg.
Multiply
Add
Add./Div./Sqrt
Div./Sqrt
Scalar Unit
Mask
Scalar Unit
Scalar
Reg.
Cache
Scalar
Exec.
2(M+A) x 4vpp x 2GHz = 16GFLOPS
Cache
Scalar
Reg.
Scalar
Exec.
2(M+A) x 2set x 4vpp x 2.2GHz = 35.2GFLOPS
Note: One vector instruction occupies one vector pipeline on SX-8R.
e.g.) The peak performance of one VFAD (vector FP add) Op. is 8.8GFLOPS
Ｐage 14
NEC
Confidential
©
NEC Corporation
2006
14
SX-8R Single Node Module
(DDR2 model)
• Up to 8 CPUs/node
CPU
- Peak Vector Performance(PVP):
35.2 GFLOPS/CPU
281.6 GFLOPS/node
CPU
....
CPU
éÂãLâØ
MM
to IXS
• Symmetric multiprocessing (SMP)
I/O
• Large Capacity Memory
- Up to 256GB
I/O ••••
I/O
Memory
• Ultra-high memory bandwidth
- 70.4GB/s per CPU
- Total 563.2GB/s per node
• Large I/O throughput
- 12.8GB/s per node
CPU
Node Module
Ｐage 15
NEC
Confidential
©
NEC Corporation
2006
15
Large Scale Multi Node System
High
Highspeed
speedprocessing
processingof
oflarge
largedata
datawith
withhigh
highperformance
performancesingle
singlenode,
node,
large
largenumber
numberof
ofnodes,
nodes,and
andhigh
highspeed
speedinterconnects
interconnectsamong
amongnodes
nodes
Key points for
high performance
1. Single node performance Max 281.6GFLOPS
2. Maximum number of node Max 512 nodes
Very efficient
3. Data transfer rate among Max 8TB/s
non-blocking switch
（Peak data transfer rate）
nodes
High speed inter-node switch (IXS)
16GB/s x 2 / nodes
ノード
最大8CPU
ノード
ノード
最大8CPU
Max
8CPU 281.6GFLOPS
最大8CPU
ノード
ノード
最大8CPU
最大8CPU
最大8CPU
最大8CPU
281.6GFLOPS
ノード
Maxノード
8CPUノード
最大8CPU
#0C
#0
#0
C
C
#0C #0C
#0
CC
CC
C C #0
C CCCC
CC
C C
C
C #0
CC
....
....P P
........
Interconnection
P
....
P ....
CPUP CPUP .... CPUP
PPPP
P ....
PP ....
PPPP....
CPU
PP Optical
P
PP
CPUP CPU
U
U
U
U UUUU
UU U UUUU
U U
U
UUU
U UU
MMU
MMU
éÂãLâØ
éÂãLâØ
MMU MMU
MMU
MMU
MMU éÂãLâØ
éÂãLâØ
MMU éÂãLâØ
MMU
éÂãLâØ
MMU
éÂãLâØ
éÂãLâØ
éÂãLâØ
éÂãLâØ
IOF
IOF
IOF
IOF
IOF
IOF
IOF
IOFIOF
IOF
IOF
IOFIOF
IOF
IOF
IOF
IOF
IOF
IOF
IOF
IOF
IOFIOF
IOF
IOF
IOF
IOF
IOF
IOF
IOF
#6 #7512
Node#0 #1 #2 #3 #4 #5 Max
Ｐage 16
NEC
Confidential
©
NEC Corporation
2006
IOFIOF
IOF
IOFIOF
IOFIOF
nodes
Node#511
16
IPF Server Roadmap
Node
Performance
CY2001
CY2001
2002
2002
Itanium
Itanium2
2003
2003
2004
2004
2006
2006
Montecito
(peak GFLOPS)
500
Itanium2 6M
Madison 9M
400
300
200
July 2001
TX7/AzusA
Announcement
51.2GFLOPS
/16CPU
192GFLOPS
/32CPU
July 2002
TX7/i9000
release
AsAmA2
128GFLOPS
/32CPU
409.6GFLOPS
204.8GFLOPS
/64CPU
/32CPU
100
50
0
Ｐage 17
NEC
Confidential
©
NEC Corporation
2006
17
Platform for PC Cluster
Enable
Enableto
tochoose
chooseOptimized
Optimizedserver
server
depending
dependingon
onusage
usageand
andbudget
budget
Opteron/Xeon Rack
Express5800/100 series, Opteron server
Cost performance Opteron/Xeon server
Express5800/T220Rc-1(Opteron)
Express5800/T120Rb-1（Xeon）
Xeon Blade
Good
Goodprice/performance
price/performanceratio.
ratio.
More
memory
slots
achieve
More memory slots achieve
system
systemcost
costreduction.
reduction.
Express5800/BladeServer
High density & Cost performance Xeon blade
Express5800/120Ba-4
Ｐage 18
Higher
Higherdensity
densitythan
thancompetitors.
competitors.
PCI
slots
achieve
expandability.
PCI slots achieve expandability.
NEC
Confidential
©
NEC Corporation
2006
18
Contents
Ｐage 19
NEC
Confidential
©
NEC Corporation
2006
19
Processor Architecture
Scalar ALU
Vector ALU
(SIMD)
Scalar ALU
S
S
SR
SR
V
VR
Scalar Processor
Vector Processor
ADD High
Speed Engine
(Vector)
Scalar
Ｖ
Ｖ
Ｖ
Ｖ
Vector Vector
Vector Vector
Vector Vector
Vector Vector
Pentium 4
Ｐage 20
Cell
NEC
Confidential
©
NEC Corporation
2006
Ｓ
SX-8
20
Performance acceleration trend
Processor core architecture
LM
Add-on accelerator card
→One-chip Co-processor
LM
SR
LM
VR
VR
VR
LM
Next x86
CPU
Scalar
VR
VR
VR
VV
V
VRV
VR
SR
SIMD
SIMD
③Heterogeneous multi-core
VR
S
②Add-on Co-Processor
VR
①Expand SIMD units
LM
LM
LM
LM
VR
VR
VR
VR
Cell（SCE/Toshiba/IBM）
Vector
Performance
Performanceof
ofHPC
HPCProcessors
Processors
Scalar
accelerated
acceleratedby
byusing
usingSIMD/Vector
SIMD/Vector
extension
extensionapproach
approach
Ｐage 21
NEC
Confidential
©
NEC Corporation
2006
SR
Vector
Vector
VR
VR
Vector Vector
VR
VR
21
Trend of Processor Multi core Chip
HPC
HPCProcessors
Processors will
will take
take in
in Vector
Vector units
units on
on the
the CPU
CPUdie.
die.
Scalar Machine ….. Add Vector (SIMD) Function
Future
V
V
X86
S
S
S
S
V
S
Java
Java
Multi core
Cell
V
S
S
Seq
Seq
S+V Multi core
V
VVVV
SS
VVVV
SS
S+ Multi Vector
S+V Multi core
S
Vector Machine
SX
VVVV
S
VVVV
VVVV
S S
VVVV
S
VVVV
VV v V
S S
S
S+V Single core
Ｐage 22
Multi core
NEC
Confidential
©
NEC Corporation
2006
V
(S+V) Multi core
22
Core Technologies for Future Supercomputer
Hybrid System
Low Power
Consumption
Interconnect
Software
＞10,000CPU
Vector
Scalar
Special
Optical
Electorical
Selection of the
suitable platform
by applications
Power density will
go up like Nuclear
Reactor.
Realize high
sustained
performance for
various
applications
Low Power
consumption and
cooling
technology will be
required.
Ｐage 23
＞1 Petabyte
Transfer speed
・Super Parallel
between LSIs will Software
influence the
・Super large data
sustained
handling
performance
Optical
Transmission
Technology
between LSIs will
be required.
NEC
Confidential
©
NEC Corporation
2006
23
NEC continues to supply
best product to customers
through
the ceaseless quest for technology
and by
Being a quality leader in the world
Thank you !!
Ｐage 24
NEC
Confidential
©
NEC Corporation
2006
24
Ｐage 25
NEC
Confidential
©
NEC Corporation
2006
25

V - ECMWF

Transcription

Similar documents