Some details of Power 7+ Architecture Changes

Innovations in action with Power7+
IBM Bulgaria
Mikhail Saykov
[email protected]
© 2010 IBM
Corporation
IBM Internal Use Olnly
Power 795
POWER7 Portfolio
Major Features:
ƒ Modular systems with linear scalability
ƒ PowerVM Virtualization
ƒ Physical and Virtual Management
ƒ Roadmap to Continuous Availability
ƒ Binary Compatibility
ƒ Energy / Thermal Management
Power 780
Power 770
Dual Socket
Quad Socket
Power 750
Power
720 / 740
Power
710 / 730
Power 775
Flex
2S/4S
Power 755
Software
BladeCenter
PS700 / PS701 / PS702
PS703 / PS704
AIX 7.1
IBM i 7.1
© 2012 IBM Corporation
Power Processor Technology Roadmap
POWER4/4+
180/130 nm
POWER5/5+
130/90 nm
9Dual Core
9Enhanced Scaling
9SMT
9Distributed Switch +
9Dual Core
9Chip Multi Processing 9Core Parallelism +
9FP Performance +
9Distributed Switch
9Memory Bandwidth +
9Shared L2
9Virtualization
9Dynamic LPARs (32)
2001
2004
POWER6/6+
65/65 nm
9Dual Core
9High Frequencies
9Virtualization +
9Memory Subsystem +
9Altivec
9Instruction Retry
9Dynamic Energy Mgmt
9SMT +
9Protection Keys
2007
POWER7/7+
45/32 nm
9Eight Cores
9On-Chip eDRAM
9Power-Optimized Cores
9Memory Subsystem ++
9SMT++
9Reliability +
9VSM & VSX
9Protection Keys+
2010
Future
© 2012 IBM Corporation
Processor designs
POWER5
POWER5+
POWER6
POWER7
POWER7+
130nm
90nm
65nm
45nm
32nm
Size
389 mm2
3245 mm2
341 mm2
567 mm2
567 mm2
Frequencies
1.65 GHz
1.9 GHz
4-5 GHz
3-4 GHz
3.6-4.4 GHz
L2 Cache
1,9MB
Shared
1,9MB
Shared
4 MB / Core
256 KB
per core
256 KB
per Core
L3 Cache
36 MB
35 MB
32 MB
4 MB / Core
10 MB / Core
Out-of-Order
Out-of-Order
In-of-Order
Out-of-Order
Out-of-Order
10 / Core
10 / Core
10 / Core
10 / Core
20 / Core
Technology
Architecture
LPAR
4
© 2012 IBM Corporation
POWER7+
POWER7
45 nm
POWER7+
POWER7
32 nm
© 2012 IBM Corporation
POWER7+
POWER7
45 nm
POWER7+
POWER7
32 nm
Add additional Cache
© 2012 IBM Corporation
POWER7+
POWER7
45 nm
POWER7+
POWER7
32 nm
Add additional Cache
Add on Chip Accelerators
© 2012 IBM Corporation
POWER7+ Design
Physical Design:
• 8 cores with integrated Cache, Memory
Controllers, and Accelerators
• 3 / 4 / 6 / 8 Core options
• 32nm technology
Features:
• 2.5X increase in L3 Cache
•eDRAM technology
•Higher Frequencies
• Memory Compression Engine
•Active Memory Expansion with no processor
overhead penalty
• Encryption / Cryptography Support
• Random Number Generator
• Enhanced Energy / Power Gating
• 1/20 LPAR Core Granularity
• 2X SPFP performance
8
SMP Fabric
Core
Core
L2
L2
L3 Cache
M
C
A
c
c
E
n
g
Core
Core
L2
L2
L3 Cache
M
C
Power Bus
L3 Cache
L2
Core
L2
Core
G
X
B
u
s
L3 Cache
L2
L2
Core
Core
SMP Fabric
© 2012 IBM Corporation
Improved SP Floating Point performance
POWER7/7+ has four FP pipelines.
Each pipeline in POWER7/7+ can do either a SP or a DP.
DP width is 2x of SP.
POWER7/7+ feeds two SP in each DP pipeline
ƒ Needed separate area for control and some status bits.
In POWER7: Two DP pipe together can execute a 2-way SIMP FPDP
instruction.
In POWER7+: Same two DP pipe together now can execute a 4-way SIMD
SPFP instruction.
Two 4-way SIMD SPFP instruction can be executed in POWER7+
ƒ Only one in POWER7.
Multiply-add counting as two, POWER7+ has 16 SP FLOP per cycle
9
© 2012 IBM Corporation
POWER7+ Encryption Accelerator
Technology
Client Benefits
ƒCrypto Offload Accelerators:
o Provide cryptographic
engines to relieve the P7+
processor from the performance
intensive cryptographic
algorithms of AES, SHA, and
RSA.
ƒCan be applied to a broader set of data
creating a stronger security ecosystem
ƒPOWER7+ core focused on application
performance.
ƒTwo primary AIX security applications:
1) Encrypted File Systems protecting
your data in storage or on backup media
2) IPsec protecting your data over the
network.
ƒHigh quality random
numbers:
oGenerated with high
performance with the RNG
offload feature of the P7+
processor.
ƒEnsures that the high demand for
entropy on a heavily loaded systems
always yields high quality random
numbers.
ƒRNG offload provides entropy snd
security is ensured
ƒProcessor performance focused on your
applications.
10
© 2012 IBM Corporation
Benefits to POWER7+ AME accelerator…
POWER7 Act Mem
Exp CPU Time
Work of
compress /
decompress
CPU TIME
Work of
managing
pages
POWER7+ Act Mem
Exp CPU Time
Work of compress / decompress
Work of
managing
pages
•Less CPU for the
same amount of
memory expansion
•Can then run more
partitions or work per
partition
•If fewer cores needed,
may result in lower
software licensing
•OR more memory
expansion for the
same amount of
processor
•Better able to relieve
memory shortages and
improve performance
•May be able to do more
work
Work done with software
Work done by hardware accelerator
© 2012 IBM Corporation
New Power7+ 770/780
12
© 2012 IBM Corporation
POWER7+ 770
✓POWER7+
✓Frequencies:
ƒ 4C SCM @ 3.8 GHz
Max Config:64 Cores
ƒ 3C SCM @ 4.2 GHz
Max Config: 48 Cores
✓Up to 64 Cores
✓Up to 4 TB of memory
✓6 PCIe Gen2 slots / CEC
✓Ethernet ports: Dual 10 Gbt & Dual 1 Gbt
✓Capacity on Demand
✓Enhanced RAS
4Socket / 4U
16-core or 12-core / 4U
13
ƒ Self-healing capability for L3 Cache functions
ƒ Core re-initialization (Running system)
ƒ Dynamic Processor Fabric Bus repair
© 2012 IBM Corporation
POWER7+ 780
✓POWER7+
✓Frequencies:
ƒ 8C SCM @ 3.7 GHz Max Cores: 128 Cores
ƒ 4C SCM @ 4.4 GHz Max Cores: 64 Cores
✓Up to 128 Cores
✓Up to 4 TB of memory
✓6 PCIe Gen2 slots / CEC
✓Ethernet ports: Dual 10 Gbt & Dual 1 Gbt
✓Enhanced Capacity on Demand options
✓Enhanced RAS
4Socket / 4U
32-core or 16-core / 4U
14
ƒ Self-healing capability for L3 Cache functions
ƒ Core re-initialization (Running system)
ƒ Dynamic Processor Fabric Bus repair
© 2012 IBM Corporation
Power 770(MMC) / 780(MHC) Diagram…..
Inter Node Buses
DRAM
DRAM
DRAM
DRAM
8 SN
Dimms
DRAM
DRAM
DRAM
DRAM
SN
SN
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
X Bus
Y Bus
Z Bus
P7
P7
A/B
Buse
s
DRAM
DRAM
DRAM
DRAM
8 SN
Dimms
DRAM
DRAM
DRAM
DRAM
SN
SN
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
CPU CARD
FSP
Card
A/B
Buse
s
P7IOC
Hub
PSI
TPMD
GX++
Busses
GX++
Busses
I/O
Backplane
P7IOC
Hub
PCIe
PSI
PCIe
12X IB
12X IB
Write Cache
EXP
SAS
SAS
2x
HMC
GX++
GX++
EXP
Write Cache
SATA
Media
DASD Backplane
Serial
USB
Gen2
Slots
15
P P P P P P
C C C C C C
I I I I I I
e e e e e e
Anchor
Card
Serial USB
Ext
SAS
RAID RAID
Battery Battery
Ethernet
2x 1Gbt & 2x 10Gbt
© 2012 IBM Corporation
POWER7+ 770 / 780 Quad Socket Planar…..
TPMD
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
SN
SN
SN
SN
SN
SN
SN
SN
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
SN
SN
SN
SN
SN
SN
SN
SN
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
DRAM
P7+
P7+
P7+
P7+
PSI
PSI
GX++
Busses
GX++
Busses
FSP
Card
P7IOC
Hub
P7IOC
Hub
I/O
Backplane
PCIe
PCIe
2x
HMC
CPU
CARD
GX++
GX++
12X IB
PCIe x8
12X IB
PCIe x8
Write Cache
SAS
EXP
SAS
EXP
Write Cache
PCIe Buses
PCIe
Buses
SATA
USB
P P P P P P
C C C C C C Anchor
Card
I I I I I I
e e e e e e
USB
Ethernet
Media
DASD Backplane
Serial
Seria
l
Ext
SAS
RAID RAID
Battery Battery
2 x 1Gbt & 2 x 10Gbt
16
© 2012 IBM Corporation
POWER7+ System Upgrades….
Power 570 & 770 systems can upgrade to POWER7+
POWER7 780
9179-MHC
POWER7 780
POWER7+ 780
9179-MHB
9179-MHD
POWER6 570
9117-MMA
POWER7 770
9117-MMC
POWER7 770
POWER7+ 770
9117-MMB
9117-MMD
POWER6 570
9117-MMA
9406-MMA converted to 9117-MMA
17
© 2012 IBM Corporation
795 Hardware Oct Announcements
2X Max Memory
Now up to 16 TB
PCIe Gen2
Two new GX++ Adapters
16Gb Fibre Channel
10Gb FCoE / CNA
RoCE support
18
© 2012 IBM Corporation
Thai
Russian
‫תודה‬
Traditional Chinese
Hindi
Hebrew
Siyabonga
Dankie
Obrigado
Zulu
Brazilian Portuguese
Arabic
Danke
Gracias
German
Afrikaans
Merci
Spanish
French
Japanese
Tamil
Korean
Grazie
Italian
Simplified Chinese
19
© 2012 IBM Corporation