Professional Documents
Culture Documents
SPARC @ Oracle
5 Processors in 4 Years
2010
2011
2012
2012
2013
SPARC T3
SPARC T4
SPARC T5
SPARC M5
SPARC M6
16 S2 cores
4MB L3$
40 nm technology
1.65 GHz
8 S3 Cores
4MB L3$
40nm Technology
3.0 GHz
16 S3 Cores
8MB L3$
28nm Technology
3.6 GHz
6 S3 Cores
48MB L3 $
28nm Technology
3.6 GHz
12 S3 Cores
48MB L3$
28nm Technology
3.6 GHz
3
Extreme
Performance
Computing
Efficiency
Optimized for
Oracle Software
Hardware and Software
Engineered, Tested, and
Supported Together
Deploy Technology Faster, with
Less Risk and Lower Costs
http://blog.oracle.com/bestperf
Copyright 2014 Oracle and/or its affiliates. All rights reserved. |
M7 Processor
32 SPARC Cores
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
DDR4 DRAM
CORE
CLUSTER
ACCELERATORS
MEMORY CONTROL
ACCELERATORS
MEMORY CONTROL
CORE
CLUSTER
Extreme
Performance
CORE
CLUSTER
M7 Core (S4)
Extreme
Performance
M7 Core Cluster
CORE
CORE
256KB L2-I$
L2-D$
L2-I$
L2-D$
CORE
CORE
70
64
64
140
L3$
PARTITION
64
140
70
64
L3$
PARTITION
SMP
&
I/O
64
70
64
64
64
64
70
140
L3$
PARTITION
64
64
256
256
ON-CHIP NETWORK
256
256
64
64
L3$
PARTITION
64
64
L3$
PARTITION
64
64
64
SMP
&
I/O
64
64
L3$
PARTITION
CORE
CLUSTER
64
64
L3$
PARTITION
MEMORY &
ACCELERATORS
MEMORY &
ACCELERATORS
L3$
PARTITION
CORE
CLUSTER
CORE
CLUSTER
140
70
CORE
CLUSTER
140
70
CORE
CLUSTER
140
70
CORE
CLUSTER
140
70
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
L3$
L3$
PARTITION PARTITION
L3$
L3$
PARTITION PARTITION
ON-CHIP NETWORK
L3$
L3$
PARTITION PARTITION
L3$
L3$
PARTITION PARTITION
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
L3$
L3$
PARTITION PARTITION
L3$
L3$
PARTITION PARTITION
ON-CHIP NETWORK
L3$
L3$
PARTITION PARTITION
L3$
L3$
PARTITION PARTITION
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
CORE
CLUSTER
10
PCIe Gen3
Interfaces
DDR4
Interfaces
DDR4
Interfaces
16 DDR4-2133/2400/2667 Channels
Very Large Memory, Up to 2TB per Processor
160GB/s (DDR4-2133) Measured Memory Bandwidth
(2X to 3X Previous Generations, T5 and M6)
DIMM Retirement Without System Stoppage
PCIe Gen3
11
M7 Processor Performance
4
3.5
3
2.5
2
1.5
1
0.5
0
M6 Baseline
Memory BW
Int Throughput
OLTP
Java
ERP
FP Throughput
12
64Bytes
version
64Bytes
version
64Bytes
version
64Bytes
version
64Bytes
version
64Bytes
version
64Bytes
version
64Bytes
Version Memory
Metadata Data
version
address
Version Miscompare
ld
st
version
address
Reference
Versions
13
New Memory
Region
Relocating
Objects
Concurrent
Thread
Accesses
Relocating
Objects
14
Embedded Object
Metadata
(ADI, Migration)
Caches Object State Table Information into Object Pointer (Pointer Coloring)
Eliminates De-reference and Memory Load from Critical Path
Copyright 2014 Oracle and/or its affiliates. All rights reserved. |
15
DB
MEMORY or L3$
Column Format
Compressed
DB
Up to 32 Concurrent DB Streams
M7 In-Silicon
Query
Engines
16
Local SRAM
Data
Input
Queues
On-Chip
Network
Data
Input
Queues
Decompress
Unpack/
Alignment
Predicate
Eval
Result
Format/Encode
Decompress
Unpack/
Alignment
Predicate
Eval
Result
Format/Encode
Decompress
Unpack/
Alignment
Predicate
Eval
Result
Format/Encode
Decompress
Unpack/
Alignment
Predicate
Eval
Result
Format/Encode
Data
Output
Queues
On-Chip
Network
Data
Output
Queues
17
Prepare
Inputs
Core Thread
Read
Results
Initiating
Call
HW Accelerator
Post
Completion
Status
Time
18
10
9
8
5
4
2
1
0
T5 Baseline
(1 Thread)
M7
(1 Query Pipeline)
19
20
M7 SMP Scalability
>1 TB/s
21
M7 SMP Scalability
5.2
TB/s
Up to 1K Cores, 8K Threads
Up to 64TB Memory
22
Server Node 1
Server Node 2
Memory
Memory
Memory
Non-Cache Segment
Non-Cache Segment
Non-Cache Segment
Remote Segment
Node 2
C
Remote Segment
Node 2
C
Remote Segment
Node 0
A
X
Remote Segment
Node 1
B
Remote Segment
Node 0
A
Remote Segment
Node 1
B
Home Segment
A A'
Home Segment
B
Home Segment
B
C
XA
Cores/Caches
A A'
B
Cores/Caches
C
B
C
A
X
Cores/Caches
Cluster-wide Security
23
M7 Server Nodes
Up to 64 Processor Cluster
Cluster
Switch
1.3
TB/s
M7 Server Nodes
Copyright
2014 Oracle and/or its affiliates. All rights reserved. |
24
M7 Summary
Extreme
Performance
Significant Increase in Processor
Performance
Further Increase Core and
Thread Performance
Increased Bandwidths Across
Caches, Memory, Interconnects
and I/O
Very Large Memory
Computing
Efficiency
Increased Virtualization Density
Low Latency Application
Migration
Flexible Logical and Physical
Partitioning
Fine-grain Power Management
Optimized for
Oracle Software
Improved Security and
Reliability via Real-time
Application Data Integrity
Concurrent Object Migration
and Pointer Coloring
Database In-Memory
Columnar Decompression,
Query Offload and Coherent
Memory Clusters
25
Acronyms
27