AMD64 with Direct Connect ArchitectureA Solid Foundation for Dense Compute Environments
Jeff Underhill
HPC Business Development (CPG)
August 04
August 13, 2004 Computation Products Group - High Density Data Center Committee 2
Agenda
1. Corporate Overview
2. Dense Computing Overview (some driving factors)
3. What’s AMD’s Role In Dense Compute Environments?
4. AMD64 Product Overview & Positioning
5. What is AMD64 with “Direct Connect Architecture”?
6. Power, Performance & Density – What’s Next?
7. Examples Built On an AMD64 Technology Foundation
8. Q&A – A test that you were all listening!
9. Ed Gasiorowski – AMD Opteron™ – Power Advantage
Corporate Overview
August 13, 2004 Computation Products Group - High Density Data Center Committee 4
AMD Overview – 35 years of Innovation
Founded: 1969
Headquarters: Sunnyvale, Calif.
Employees: 14,400 worldwide
Sales Mix: >75% international
Business Units Q2 04 Revenue($M)
• Microprocessors $554
• Flash Memory $673
• Other $34
Total: $1.262 billion
2003 Revenue: $3.5 billion
Major Operations• Sunnyvale, CA • Austin, TX• Dresden, Germany• Penang, Malaysia• Aizu-Wakamatsu,
Japan• Bangkok, Thailand• Singapore• Suzhou, China
August 13, 2004 Computation Products Group - High Density Data Center Committee 5
AMD Worldwide Operations
Sunnyvale, California
• AMD Global HQ
• Submicron Development Center
• FASL LLC (Spansion™ brand) HQ
Boston, Mass.
Boston Design Center
Austin, Texas
• Fab 25 — Flash Memory
• Computations Product Group
• Personal Connectivity Solutions Group
Frimley, UK
European Service Center
Dresden, Germany
• Fab 30 — Microprocessors• Dresden Design Center
Aizu-Wakamatsu/Tokyo, Japan
• Flash Memory Fabs: JV1, 2, 3
• FASL Japan HQBangkok, Thailand
Flash Memory Assembly and Test
Suzhou, China
Flash Memory Assembly and Test
Penang/Kuala Lumpur, Malaysia
Microprocessor and Flash Assembly, Test and Packaging
Singapore
Microprocessor Assembly, Test and Packaging
August 13, 2004 Computation Products Group - High Density Data Center Committee 6
AMD Industry Leadership
AMD is a global supplier of integrated circuits for the personal and enterprise computing, communications and
consumer electronic markets
1. CPG - Leading the industry’s transition to pervasive 64-bit computing with the AMD64 family of microprocessors
2. FASL - #1 in NOR Flash memory* with a goal to become the world’s overall Flash memory leader
3. PCSG - Delivering innovative embedded x86 system-on-chipsolutions for internet and wireless applications– Broadens AMD’s product portfolio and access to high-growth and
profitable emerging markets
* iSuppli Corporation, Q303 NOR Flash market share estimate
August 13, 2004 Computation Products Group - High Density Data Center Committee 7
Transistor Count / Process Migration…
16 32 64 BitsInteger FPU, SIMD Vector1/3 Issue 9 IssueTrivial Cache 1MB CacheCPU SOC
0.8µ200K
transistors48mm2
0.35µ1.2M
transistors35mm2
0.25µ9M
transistors78mm2
0.18µ37M
transistors120mm2
AMD Am386®
ProcessorAMD
Am486®
Processor
AMD K6®-III Processor
AMD Athlon™
Processor
AMD Opteron™Processor
0.13µ100M
transistors193mm2
90nm100M
transistors114mm2
AMD Opteron™Processor
August 13, 2004 Computation Products Group - High Density Data Center Committee 8
AMD Leads Pervasive 64-bit Computing
August 13, 2004 Computation Products Group - High Density Data Center Committee 9
AMD is a Product Powerhouse
2002 Samsung Best Supplier Award –AMD Flash
CeBIT Highlights 2002: Hardware Category -AMD Opteron and AMD Athlon 64
2003 INSIGHT Award for Most Innovative Memory Device – AMD MirrorBit
Cisco President's Customer
Satisfaction Award
- 2000 –AMD Flash
2004
2004
2004
August 13, 2004 Computation Products Group - High Density Data Center Committee 10
Driving AMD Brand Awareness
AMD Product Placement
in Spy Kids III
AMD Product Placement
in Spy Kids III
Dense Computing OverviewSome Driving Factors
August 13, 2004 Computation Products Group - High Density Data Center Committee 12
Dense Computing OverviewSome Driving Factors…
•Business functions are increasingly digital & distributed!
– CRM, ERP, Commerce, Communications, Inventory Management, Invoices, Risk analysis, EDA, DCC, ESD, Media Content Delivery (Akimbo, Netflix & Tivo) etc…
•We live in an increasingly digital personal world too– Electronic banking, investing, billing, shopping,
communicating, filing taxes, photography, home movies, MP3’s etc…
– All of these functions require additional resources at the various service providers!
With this migration comes an ever increasing need for storage and compute power!
August 13, 2004 Computation Products Group - High Density Data Center Committee 13
Dense Computing OverviewA case in point …
Businesses accounting is now subject to the Sarbanes-Oxley Act of 2002 stating: -
The Board must require registered public accounting firms to "prepare, and maintain for a period of not less than 7 years, audit work papers, and other information related to any audit report, in sufficient detail to support the conclusions reached in such report."
The Board must require registered public accounting firms to "prepare, and maintain for a period of not less than 7 years, audit work papers, and other information related to any audit report, in sufficient detail to support the conclusions reached in such report."
Think of the storage & retrieval implications of this single accounting act!
August 13, 2004 Computation Products Group - High Density Data Center Committee 14
Dense Computing OverviewAnother case in point … Transport Security Administration
• The TSA are considering next generation scanning technology that would allow them to dissect baggage in 3D
• They are also considering the need to keep baggage scans on file for ‘some’ period of time in order to review – should the need arise
Again - Think of the storage & retrieval implications …
August 13, 2004 Computation Products Group - High Density Data Center Committee 15
Dense Computing OverviewAnother case in point … Distribution & security
• More and more business functions require availability on a worldwide scale
• There’s a definite need to securely and reliably process and distribute ever increasing datasets
• This translates into a direct need for large memory models and increased compute power to execute complex encryption & compression algorithms –some at wire speed!
August 13, 2004 Computation Products Group - High Density Data Center Committee 16
RAS (Reliability, Availability, Serviceability)
•Increasing dependence on these business functions drives RAS requirements – most functions are becoming, or are already - mission critical!
•The promise of increased availability and manageability are among the incentives driving migration to dense rack and blade architectures.
•OEM’s and integrators are layering RAS features on commodity server platforms closing the feature gap on ‘big iron’ systems at a fraction of the cost.
… RAS features are essential and allow vendors to innovate and differentiate their offerings!
August 13, 2004 Computation Products Group - High Density Data Center Committee 17
Dense Computing OverviewWhat does this mean…?
In order to address increasing demands you can generally: -
Work Harder Faster ComponentsWork Smarter Process InnovationGet Help Scale Out / Parallelize
Over the past decades there have been advances in each of these areas from an electronics perspective, but now more so than ever ‘Get Help’ seems to be ‘required’ to
meet the dramatic increases ahead!
August 13, 2004 Computation Products Group - High Density Data Center Committee 18
Dense Computing OverviewIn order to ‘Get Help’ …
•More compute devices are required to cooperatively address increasingly complex and sizeable tasks.
•This in turn drives the need to fit more and more compute devices within ideally the same or at least moderately increased space.
•With increased compute devices comes increased power consumption, the increased heat generated demands additional environmental requirements (cooling, filtering etc…)
•Before too long the datacenter infrastructure could be stretched to the point that it’s a limiting factor for the required growth!
August 13, 2004 Computation Products Group - High Density Data Center Committee 19
Dense Computing OverviewMigration from rack mount to blade servers …?
•There’s a lot of activity within the computer industry to address the infrastructure issues while providing solutions that IT organizations can deploy and manage (virtualization etc…) in order to maximize density and utilization of resources.
•Virtualization of pooled resources with remote management capabilities increase the number of nodes per administrator (30:1 vs. 15:1)
•Reduce cabling by up to 87% - leading to increased reliability at lower cost.
•Improved availability with enough space to include N+1 redundancy within a blade shelf.
•Up to 24% reduction in data center floor space using blades over rack mount servers (48 vs. 30)
Source - HP ProLiant Blade Systems White Paper
August 13, 2004 Computation Products Group - High Density Data Center Committee 20
Dense Computing OverviewA quick look at the blade market …
•Blades – an explosive growth market …?
Rack 2002 Rack 2008 Blade 2002 Blade 2008
Shipments (000) 1,376 2,896 37 2,671
Dollars ($M) $8,104 $11,261 $91 $7,071
Growth Shipments (000)
110% 6987%
Growth Dollars ($M)
39% 7616%
Linux & Windows Servers – IDC Q4 2003
What’s AMD’s RoleIn Dense Compute Environments?
August 13, 2004 Computation Products Group - High Density Data Center Committee 22
AMD’s RoleIn Dense Computing Environments?
• Primarily a component vendor - however advancements in density require that the industry as a whole play their part in addressing the issues.
• CPU and Chipset components play a large role in addressing the needs for ever increasing performance at price and power points that are compelling for next generation compute devices.
• AMD has already made advances in this area and continues to innovate, enhancing the price / power / performance metric that’s so important to realizing the dense data center model.
AMD64 Product Overview & Positioning
August 13, 2004 Computation Products Group - High Density Data Center Committee 24
AMD64 – A quick show of hands …
Who’s heard of ...
1. AMD64 or x86-64 Architecture?
2. AMD Opteron™ with Direct Connect Architecture?
3. Hyper Transport?
4. Glue less multi-processing?
5. SMP / CC-NUMA architectures?
6. Intel?
August 13, 2004 Computation Products Group - High Density Data Center Committee 25
AMD64 Brand Architecture
Customer-Centric Innovation
Your Link To The Future Of Computing
Mobile Athlon64 Athlon64 Athlon64 FX OpteronAthlon32
Leading Mobile Performance
Leading Performance
AMD64 Capable
CinematicComputingon AMD64
Simplifying Business
Leading Value and Performance
x86-32
August 13, 2004 Computation Products Group - High Density Data Center Committee 26
AMD64 Solutions from Mobile to Server
AMD serves the mid-range, high-volume server space by supporting the best of both scale-up and scale-out approaches
Perf
orm
ance
Price
1P Desktop& Mobile
1P Server& Workstation
2P/4P ServersWorkstation
4P+ Servers/HPCCluster
Provide Server-class CPUs & components to build balanced, high-performance 1P to 8P Scale-Up systems;
Design these CPUs and components to fit in the cost and complexity envelope of simple Scale-Out systems;Lower pricing with “PC” volumes
Remove the complexity and price barrier between Scale-up & Scale-out
Leverage & extend the Industry investment and expertise in x86 while extending it to 64-bit.
August 13, 2004 Computation Products Group - High Density Data Center Committee 27
AMD64Innovative Approach to 64-bit Computing
• AMD is leading the way to the 64-bit future with AMD64 technology
• The first platform designed for high-end server, mainstream client, and mobile markets
• With AMD64 technology
• Preserve existing x86 32-bit investment
• Industry leading 32-bit performance
• Evolutionary ability to migrate to 64-bit
HyperTransport™ technologyIncrease overall system performance by reducing I/O bottlenecks and increasing bandwidth
Integrated DDR memory controllerImproved application performance
PowerNow!™ & Cool‘n’Quiet™ technologyPower Management Solutions to reduce power consumption & system noise
Enhanced security technologyvirus protection to prevent “buffer overflow” attacks with Microsoft® Windows® XP SP2
August 13, 2004 Computation Products Group - High Density Data Center Committee 28
AMD Athlon™ 64 Processor
Replaces Address, Data and Control Bus
72
16
AMD64: Desktop / Mobile Processor
8 Byte memory controller supporting DDR200,
DDR266, DDR333 & DDR400 Memory
Drive up to 4 unbuffered DIMMs
4 DIMMs <=DDR266 (PC2100)
2 DIMMs >=DDR333 (PC2700)
2 DIMMs DDR400 (PC3200)
Future memory support as it is defined
Up to 4GB x4 DRAMS (4GB DIMMs)
One 16-bit non-coherent HyperTransport™ I/O
link
On chip L1 & L2 cache
64KB L1 ICache, 64KB L1 DCache
512K / 1M ECC protected L2 Cache
754-pin µPGA Package
L2Cache
L1Instruction
Cache
L1Data
Cache
AMD64Processor
Core
DDR MemoryController
HyperTransport™
August 13, 2004 Computation Products Group - High Density Data Center Committee 29
1P Athlon™ 64 System
System Strengths:
Ideal for low cost, small footprint designs that still require excellent compute & I/O performance
Examples:
• Storage / network appliances
• Small form factor modular designs (1P blades)
• Application specific compute products
AMDAthlon™ 64
AMDAthlon™ 64
72b DDRw/ ECC
USB, EIDE, SATAFDC, and GbE I/O
HubI/O
Hub
PCI 33-32
KVMOpt’lBMC
Opt’lBMC SIOSIOTPMTPMBIOSBIOS
PCI-X 2.0
Opt’l SCSIDual GbE
Opt’l SCSIDual GbE
PCI-XBridgePCI-XBridge
August 13, 2004 Computation Products Group - High Density Data Center Committee 30
AMD Athlon 64 FX
L2Cache
L1Instruction
Cache
L1Data
Cache
AMD64Processor
Core
DDR MemoryController
HyperTransport™
72/144
16
High-end consumer (939 pin – 1xHT)
16 Byte memory controller supporting DDR200,
DDR266, DDR333 & DDR400 Memory
Drive up to 8 DIMMs
8 DIMMs <= DDR266 (PC2100)
4 DIMMs >= DDR333 (PC2700)
4 DIMMs DDR400 (PC3200)
Up to 4GB x4 DRAMS (4GB DIMMs)
One 16-bit non-Coherent HyperTransport™
Technology Link
On chip L1 & L2 cache
64KB L1 ICache, 64KB L1 DCache
Up to 1M ECC protected L2 Cache
939-pin µPGA Package
Replaces Address, Data and Control Bus
August 13, 2004 Computation Products Group - High Density Data Center Committee 31
1P Athlon™ 64 FX
AMDAthlon™ 64 FX
AMDAthlon™ 64 FX
128b DDRReg ECC System Strengths:
Outright client performance –Lots of compute, memory and I/O bandwidth.
Examples:
• Ultimate Gaming PC• 1P Graphics workstation• Audio / Visual content
creation
I/OHubI/O
Hub
PCI 33-32USB, EIDE, SATA
FDC, and GbE
KVMOpt’lBMC
Opt’lBMC SIOSIOTPMTPMBIOSBIOS
AGP8x
AGP8x
August 13, 2004 Computation Products Group - High Density Data Center Committee 32
1P AMD Opteron™ 100 Series
L2Cache
L1Instruction
Cache
L1Data
Cache
AMD64Processor
Core
DDR MemoryController
HyperTransport™
72/144
16 16 16
18 CAS lines for 32GB of memoryAMD64: 1 way Value Workstation, storage & networking applications
16 Byte memory controller supporting DDR200,
DDR266, DDR333 & DDR400 Memory
CHIPKILL ECC with x4 DRAMs
Drive up to 8 registered DIMMs
8 DIMMs <= DDR266 (PC2100)
4 DIMMs >= DDR333 (PC2700)
4 DIMMs DDR400 (PC3200)
Up to 4GB x4 DRAMS (4GB DIMMs)
Three 16-bit non-Coherent HyperTransport™
Technology Links
On chip L1 & L2 cache
64KB L1 ICache, 64KB L1 DCache
Up to 1M ECC protected L2 Cache
940-pin µPGA Package
Replaces Address, Data and Control Bus
August 13, 2004 Computation Products Group - High Density Data Center Committee 33
1P AMD Opteron™ 100
System Strengths:
Ideal for cost sensitive designs system where I/O and memory are critical commodities
Examples:
• Storage servers• Mid-range workstations• Blade server• Network switch / router
AMDOpteron100 Series
AMDOpteron100 Series
128b DDRReg ECC
USB, EIDE, SATAFDC, and GbE I/O
HubI/O
Hub
PCI-X 2.0
PCI 33-32
Opt’l SCSIDual GbE
Opt’l SCSIDual GbE
KVMOpt’lBMC
Opt’lBMC SIOSIOTPMTPMBIOSBIOS
AGP8x
AGP8x
PCI-XBridgePCI-XBridge
August 13, 2004 Computation Products Group - High Density Data Center Committee 34
2P - AMD Opteron™ 200 Series
72/144
16 16 16
18 CAS lines for 32GB of memoryAMD64: 2 Way Performance Workstation / Server,
high performance storage & networking.
16 Byte memory controller supporting
DDR200, DDR266, DDR333 & DDR400 Memory
CHIPKILL ECC with x4 DRAMs
Drive up to 8 registered DIMMs
8 DIMMs <= DDR266 (PC2100)
4 DIMMs >= DDR333 (PC2700)
4 DIMMs DDR400 (PC3200)
Up to 4GB x4 DRAMS (4GB DIMMs)
One coherent and two 16-bit non-
Coherent HyperTransport™ Technology Links
On chip L1 & L2 cache
64KB L1 ICache, 64KB L1 DCache
1MB ECC protected L2 Cache
940-pin µPGA Package
L2Cache
L1Instruction
Cache
L1Data
Cache
AMD64Processor
Core
DDR MemoryController
HyperTransport™
Replaces Address, Data and Control Bus
August 13, 2004 Computation Products Group - High Density Data Center Committee 35
Advanced AMD Opteron™ MP System Architecture
•Scalable memory and I/O bandwidth – Each processor adds up to 32GB of memory– Each processor adds two HyperTransport™ buses for
more PCI-X/PCI Express and other I/O bridges– Fewer chips required
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
AMDOpteronAMDAMD
OpteronOpteronAMDOpteronAMDAMD
OpteronOpteron
PCI-XBridgePCI-XBridgePCI-X
DDR DDR
AMD Opteron System
PCI-XBridgePCI-XBridgePCI-X PCI ExpressPCI Exp
BridgePCI ExpBridge
System Strengths:
Ideal for performance dense scale-out systems where node and overall system scalability are important
Examples:
• 1U rack mount server• DCC Workstation• Mid high end network /
storage processor
August 13, 2004 Computation Products Group - High Density Data Center Committee 36
4P - 8P AMD Opteron™ 800
72/144
16 16 16
AMD64: 4 - 8 Way Performance Server
16 Byte memory controller supporting
DDR200, DDR266, DDR333 & DDR400 Memory
CHIPKILL ECC with x4 DRAMs
Drive up to 8 registered DIMMs
8 DIMMs <= DDR266 (PC2100)
4 DIMMs >= DDR333 (PC2700)
4 DIMMs DDR400 (PC3200)
Up to 4GB x4 DRAMS (4GB DIMMs)
Three 16-bit Coherent HyperTransport™
Technology Links
On chip L1 & L2 cache
64KB L1 ICache, 64KB L1 DCache
1M ECC protected L2 Cache
940-pin µPGA Package
L2Cache
L1Instruction
Cache
L1Data
Cache
AMD64Processor
Core
DDR MemoryController
HyperTransport™
Replaces Address, Data and Control Bus
18 CAS lines for 32GB of memory
August 13, 2004 Computation Products Group - High Density Data Center Committee 37
Advanced AMD Opteron™ MP System Architecture
•Scalable memory and I/O bandwidth – Multi processor systems without glue logic– Each processor adds 32GB of memory– Fewer chips required
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
PCI-XBridgePCI-XBridgePCI-X
OtherBridgeOtherBridge
OtherI/O
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
DDR
DDR
DDR
DDR
AMD Opteron System
PCI ExpressPCI ExpBridge
PCI ExpBridge
System Strengths:
Ultimate performance for scale-up and scale-out systems where density, memory, compute and I/O performance are essential
Examples:
• 4U rack mount server• Ultimate HPC Cluster node• Dedicated network file
server
August 13, 2004 Computation Products Group - High Density Data Center Committee 38
Contrast to a typical MP system…
AMD Opteron System Typical MP System
• System scalability limited by MCH ports– Maximum of 4 processors
o Processors compete for FSB bandwidth with memory and I/O accesses
– Memory size and bandwidth are limited– Limited number of bridges– Additional chips required
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
PCI-XBridgePCI-XBridgePCI-X
OtherBridgeOtherBridge
OtherI/O
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
DDR
DDR
DDR
DDR
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
PCI-XBridgePCI-XBridge
MemoryCtlr Hub(MCH)
MemoryCtlr Hub(MCH)
ProcessorProcessorProcessor
ProcessorProcessorProcessor
ProcessorProcessorProcessor
ProcessorProcessorProcessor
PCI-X
PCI-XBridgePCI-XBridge
PCI-XBridgePCI-XBridge
DDR
DDR
MemoryExpanderMemory
Expander
MemoryExpanderMemory
Expander
PCI-X
PCI ExpBridge
PCI ExpBridge
PCI-XPCI Express
AMD Opteron™ with Direct Connect Architecture?Hyper Transport?Glue less multi-processing?SMP / CC-NUMA architectures?
August 13, 2004 Computation Products Group - High Density Data Center Committee 39
Advanced AMD Opteron™ 8P System Architecture
•Scalable memory and I/O bandwidth – Up to 8 processors without glue logic– Each processor adds 32GB of memory– Fewer chips required
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
PCI-XBridgePCI-XBridgePCI-X
OtherBridgeOtherBridge
OtherI/O
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
DDR
DDR
DDR
DDR
AMD Opteron System
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
AMDOpteronAMDAMD
OpteronOpteron
DDR
DDR
DDR
DDR
PCI ExpressPCI ExpBridge
PCI ExpBridge
System Strengths:
Ultimate performance for dense scale-up and scale-out systems where memory, compute and I/O performance are essential
Enough memory (256GB), compute and I/O bandwidth to accommodate even the most intense applications
August 13, 2004 Computation Products Group - High Density Data Center Committee 40
AMD Athlon™ 64 and Opteron™ FeatureEnhanced Virus Protection with AMD64
VIRUS ENHANCED ENHANCED VIRUS VIRUS
PROTECTIONPROTECTIONX+
Upcoming
Enhanced Virus Protection–Helps prevent “buffer overflow” virus attacks from harming the PC
and spreading through the network• Activated in 32-bit Windows® XP with Service Pack 2 – Just Released!• Designed to be native to 64-bit Windows XP and Windows “Longhorn”
client OS
Service Pack 2
August 13, 2004 Computation Products Group - High Density Data Center Committee 41
Register Extension
• AMD6464-bit integer registers
40-bit Physical Address
48-bit Virtual Address
• Register ExtensionsSixteen 64-bit integer registers
Sixteen 128-bit SSE registers
• Vector Math Instruction Set3DNow!™
SSE & SSE2 support
• Done in such a way as to not alter the IA32 instruction set.
RAX
63
GGPPRR
xx8877
079
31
AHEAX AL
0715In x86
XMM0SSSSEE
127 0
XMM7
EIP
EDI
Added by x86-64
EAX
XMM8XMM8
XMM15
R8
R15
http://www.x86-64.org
August 13, 2004 Computation Products Group - High Density Data Center Committee 42
AMD’s Solution is an Evolutionary Approach:Backward compatible to IA32
AMD: Single PlatformAMD: Single Platform Designed to maintain legacy Designed to maintain legacy compatibility …compatibility …
…… and obey the Immutable Laws.and obey the Immutable Laws.
(< 3GB Limit) (4GB Limit)Leverage existing infrastructure – thermal, enclosures, power, and BIOS
Run existing 32-bit applications natively
Allow 64-bit migration according to your schedule
Low learning curve for users and support staff
August 13, 2004 Computation Products Group - High Density Data Center Committee 43
AMD OpteronTM Software A Word on Architectural Feedback
• Code size is only up about 5%– Mostly due to 64-bit literals
• Instruction count is down about 15%– Additional registers really paying off– Many spill/fill memory references eliminated– Call-Exit sequences vastly improved
• Reduced instruction count and increased Instructions Per Cycle (IPC) mean substantial performance gains
– AMD Opteron™ IPC improves about 5%– AMD64 instruction count down about 15%– Net improvement about 20%– Your mileage will vary
AMD64 With Direct Connect Architecture
August 13, 2004 Computation Products Group - High Density Data Center Committee 45
Advanced AMD Opteron™ Direct Connect Architecture
The AMD64 family of products leverage the low latency high bandwidth HyperTransport bus to
achieve a distributed architecture
August 13, 2004 Computation Products Group - High Density Data Center Committee 46
HyperTransport™ Technology Consortium
August 13, 2004 Computation Products Group - High Density Data Center Committee 47
Advanced AMD Opteron™ Direct Connect Architecture
The AMD Opteron™ provides three 16-bit HyperTransportlinks providing 6.4GB/s bandwidth per link + dual
channel DDR400 memory interfaces!
800MHz * 2(DDR) * 16(width) * 2(Bi-directional) = 51.2Gb/s /8 = 6.4GB/s
2 x PC3200 DDR interface = 6.4GB memory bandwidth
3 x HT links = 19.2GB/s Total I/O bandwidth
+ 6.4GB/s (memory bandwidth) = 25.6GB/s total CPU bandwidth!
and Increasing!
August 13, 2004 Computation Products Group - High Density Data Center Committee 48
Advanced AMD Opteron™ Direct Connect Architecture
AMD Opteron System
AMDOpteron™
•Scalable memory and I/O bandwidth – Up to 8 processors without glue logic– Each processor adds 32GB of memory– Fewer chips required– Each HT link provides 6.4GB/s of I/O bandwidth
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
PCI-XBridgePCI-XBridgePCI-X
OtherBridgeOtherBridge
OtherI/O
DDR
DDR
DDR
DDR
PCI ExpressPCI ExpBridge
PCI ExpBridge
–CPUs are connected directly to CPUs for more linear symmetrical multiprocessing
–Memory is connected directly to the CPU optimizing memory performance
–I/O is directly connected to the CPU for balanced throughput and scaleable I/O
Addresses and helps eliminate the real
challenges and bottlenecks of a complete system
architecture
AMDOpteron™
AMDOpteron™
AMDOpteron™
August 13, 2004 Computation Products Group - High Density Data Center Committee 49
Advanced AMD Opteron™ Direct Connect Architecture
Ease of design & layout
AMDOpteron™
I/OHubI/O
Hub PCIIDE, FDC,USB, Etc.
PCI-XBridgePCI-XBridgePCI-X
OtherBridgeOtherBridge
OtherI/O
DDR
DDR
DDR
DDR
PCI ExpressPCI ExpBridge
AMDOpteron™
AMDOpteron™
AMDOpteron™
AMDOpteron™
PCI ExpBridge
- 1P done in 4 layers 4.7mil nominal (FR4 2116 ply) dielectric- 2P in as little as 6 layers!
Power, Performance & DensityWhat’s Next?
August 13, 2004 Computation Products Group - High Density Data Center Committee 51
Low Power AMD Opteron™ Processors Overview
•Lower wattage processors are expected to enable dense platforms where lower power requirements are essential.
•AMD launched low power parts in February 2004. This is enabling vendors to effectively service additional market segments:
–Standard servers and Blades: In rack-mount chassis, lower power is needed to enable 1U/2P, 1U/4P, 2U/4P and other dense designs to meet end-user needs for overall lower power consumption
–Embedded controllers: In networking, mass storage, telecommunications and other markets low power is needed:
• Network attached storage (NAS) market needs full microprocessor capability with less power than standard parts consume
• Storage area networks (SAN) devices demand very high I/O bandwidth but cannot make provisions for custom thermal solutions
August 13, 2004 Computation Products Group - High Density Data Center Committee 52
Low Power AMD Opteron™ Processors Overview (con’t)
• These low power parts are bringing AMD64 technology to a broad range of server, workstation and non-traditional enterprise products.
• These new products extend the current AMD Opteron product line– The full-power CPU roadmap is 89 watts 95 watts– The mid-power CPU roadmap is 55 watts– The low-power CPU roadmap is 30 watts
• There is no performance penalty when using low power parts.– Low power parts perform exactly to equivalent speed grade full power
parts
• Lower power AMD Opteron processors leverage today’s current infrastructure.
– Continuing the architecture (940 pin socket)– 1xx, 2xx & 8xx parts are available in 55watt and 30watt today!– Leverages existing infrastructure: MB, Cooling, Software
August 13, 2004 Computation Products Group - High Density Data Center Committee 53
Low Power Product Positioning
•Rack Dense and Blade Servers (55 watt)–Where performance per watt is a key criteria–Designed for high-density without sacrificing high-
performance–Many end-user computing facilities cannot support full racks of
full power processors
•Modular Environments (30 watt)–Where I/O’s per watt is a key criteria–3 HyperTransport™ links enable substantial I/O for demanding
markets–The AMD Opteron feature set is a perfect fit for non-standard
servers, networking, storage and appliances
•One architecture for all AMD Opteron implementations–89w, 55w and 30w processors all leverage existing 940
infrastructure• Layout design, thermal solutions, software, etc.
August 13, 2004 Computation Products Group - High Density Data Center Committee 54
AMD Opteron Processor™ HE (55 watt):Leading Performance Per Watt
SPECcpu®2000 Performance
1335 13391354 1394
0
400
800
1200
1600
SPEC®int_peak2000 SPEC®fp_peak2000
SPECcpu®2000 Performance/Watt
15.0 15.0
24.6 25.3
0
10
20
30
SPEC®int_peak2000/Watt SPEC®fp_peak2000/Watt
AMD Opteron™ Model 146 (89W)
AMD Opteron Model 146 HE (55W)
The AMD Opteron Processor Model 146 HE provides equivalent performance at almost 40% lower wattage compared to the full power AMD Opteron
Processor Model 146
August 13, 2004 Computation Products Group - High Density Data Center Committee 55
Typical Server Architecture
Fire
wal
lFi
rew
all
Cache Cache serversservers
Load Load Balancing Balancing
serversserversWeb Web
serversservers(1(1--2P)2P)
Java Java Application Application
serversservers(2(2--4P)4P)
Database Database serversservers(4(4--8P)8P)
StorageStorage
Tier 1Tier 1 Tier 2Tier 2 Tier 3Tier 3Basic web contentBasic web content Transactional content Transactional content
ClientsClients
Inte
rnet
/
Inte
rnet
/
Intr
anet
Intr
anet
Enterprise View
File & Print File & Print serverserver(1(1--2P)2P)
NetworkNetwork Enterprise KeySMB View
Clie
nts
Clie
ntsClientsClients
InternetInternet PrinterPrinter
August 13, 2004 Computation Products Group - High Density Data Center Committee 56
AMD64 architecture performance?
Application-specific benchmarks demonstrate the value of the AMD Opteron™ processor in:
–File and print – Netbench
–Web infrastructure – HP’s published WebBench™ 5.0 benchmark shows one AMD Opteron processor can outperform two Xeon processors!
–Web security - SpecWeb99_SSL
–Data Warehousing – TPC-H demonstrates an AMD Opteron processor-based cluster outperforms a Xeon processor-based cluster costing three timesmore!!!
August 13, 2004 Computation Products Group - High Density Data Center Committee 57
Veritest File Serving Performance …
April 2004 - AMD commissioned VeriTest to conduct a series of tests comparing File Serving performance of AMD and Intel-based
servers running Windows 2003 Server, Standard Edition.
August 13, 2004 Computation Products Group - High Density Data Center Committee 58
AMD Opteron™ ProcessorFile & Print Serving Performance
•File and print performance:–Heavily I/O dependent
•AMD Opteron™ edge:–HyperTransport™ technology delivers superior I/O performance to make AMD Opteron processor one of the best performing file and print servers
•Key take away:–In the dual processor test, AMD Opteron 248 CPUs outperformed Xeon 3.2GHz by 35%
–For price sensitive file server needs, a single AMD Opteron processor is competitive with a dual Xeon configuration
NetBench 7.02
Samba Server File Sharing (Linux v2.4.23 kernel)
MB/s
41.12
38.86
49.37
Itanium 2 1.5GHz
Xeon 3.06GHz
AMD Opteron 248
100%
106%
127%64-bit
32-bit
64-bit
™
InfoWorld - January 30, 2004
August 13, 2004 Computation Products Group - High Density Data Center Committee 59
Clie
nts
Clie
nts
AMD Opteron™ ProcessorWeb Infrastructure Performance
WebBench 5.0(Windows Server 2003 Standard, 1-2P Servers)
0500
10001500200025003000350040004500
1 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60
Number of Clients
Req
uest
s P
er S
econ
d
AMD Opteron 248 2P AMD Opteron 248 1PXeon 3.2GHz 2P Xeon 3.2GHz 1P
Hewlett-Packard Results
All scores 32-bit
AMD Opteron 2P wins by 57%
™
HP ProLiant DL140 (Xeon) vs. DL145 (AMD Opteron)
1P AMD Opteron processor can outperform 2P Xeon by over 20%
2P AMD Opteron processor can outperform 2P Xeon by 57%
• Web server performance: – Highly dependent on memory
response and I/O performance
• AMD Opteron™ processor edge:– HyperTransport™ technology and
the integrated memory controller of AMD Opteron processors help to deliver great I/O performance and memory response time in handling large numbers of user connections and response times both on static and dynamic web pages.
• Key take away:– AMD Opteron processor’s Direct
Connect Architecture is perfect for web serving, delivering industry-leading performance that is sustained as the client load increases.
August 13, 2004 Computation Products Group - High Density Data Center Committee 60
Clie
nts
Clie
nts
AMD Opteron™ ProcessorWeb Security Performance
• Web security performance:– Highly dependent on memory latency– As encryption gets more complex, 64-bit
processing will become necessary
• AMD Opteron™ edge: – The integrated memory controller of
AMD Opteron processor enhances the caching of secure data
– RSA Software is optimized for the AMD Opteron processor to enable fast transaction response
– AMD64™ technology enables seamless transition to 64-bit computing as encryption algorithms increase in complexity
• Key take away: – Web security requirements will drive a
migration to more complex algorithms best served by 64-bit technology
– The AMD Opteron processor outperforms Xeon in today’s 32-bit secure Web serving environment and helps make the transition to 64-bit computing easy
1304138214401475
19302000
2340
Xeon 3.06GHz
Xeon 3.06GHz 1MB L3
Xeon 3.2GHz 1MB L3
Xeon 3.2GHz 2MB L3
Itanium 2 1.5GHz 6MB L3
AMD Opteron 246
AMD Opteron™ 248 117%
100%
72%
69%
65%
97%
74%
SPECweb®99_SSL(2P Servers)
AMD Opteron™ and Itanium 2 scores 64-bit
Xeon scores 32-bit
August 13, 2004 Computation Products Group - High Density Data Center Committee 61
AMD Opteron™ ProcessorApplication Server
• Application Server Performance– Highly dependent on memory
response and I/O sub system performance
• AMD Opteron™ processor edge:– HyperTransport™ technology and the
integrated memory controller of AMD Opteron processor delivers fantastic I/O performance and memory response time in handling large numbers of user connections and response times both on static and dynamic Web pages.
• Key take away:– AMD Opteron processors are cost-
effective for application servers– AMD Opteron processor’s Direct
Connect Architecture is able to drive high application server performance TOPS: Total Operations Per Second
August 13, 2004 Computation Products Group - High Density Data Center Committee 62
Clie
nts
Clie
nts
AMD Opteron™ ProcessorData Warehousing Performance
•Data Warehousing performance:
–Highly dependent on server scalability, efficiency and data throughput capacity
•AMD Opteron™ edge:–Direct Connect Architecture offers
scalable memory bandwidth with very low latency for better data throughput as data requirements increase
•Key take away:–Datasets are increasing at huge
rates requiring more efficient data warehousing systems
–AMD Opteron processor-based clusters can outperform a Xeon processor-based cluster at one third the cost!
12216AMD Opteron™ 246
12995
13194
Pentium® III Xeon 900MHz (64CPUs)
AMD Opteron™ 246 (16 CPUs)
TPC-H™(clustered 2P Servers)
100GB Database
300GB Database
100%
102%$863,410
$2,636,750
AMD Opteron scores 64-bit
Xeon score 32-bit
August 13, 2004 Computation Products Group - High Density Data Center Committee 63
AMD Opteron™ Dual Core
AMD Announces Technology Milestone With Its Multiple-Core Strategy
— First AMD64 dual-core solutions planned to be available in mid-2005 —
SUNNYVALE, CALIF. -- June 14, 2004 --Today AMD (NYSE: AMD) announced a technology milestone with the completed design of its AMD64 dual-core processors. AMD plans to deliver high-performance dual-core products to the x86 server market in mid-2005 and introduce dual-core solutions for high-end client PCs in the second half of 2005.
August 13, 2004 Computation Products Group - High Density Data Center Committee 64
AMD Opteron™ Dual Core Highlights
• Dual core parts will launch in the AMD Opteron processor line in 2005 and will enable AMD to continue to drive technology leadership in the server and workstation market.
–These processors can be inserted into 90nm AMD Opteron 940-pin sockets with a BIOS update, saving time and costs while increasing performance and value.
–Completely compatible with x86 and AMD64 applications while benefiting multi-threaded environments.
–CPU intensive applications will experience the greatest performance advantage. Examples include: Business Processing, Technical Computing, Network processing and Web Infrastructure.
–AMD Opteron with dual core processing is another example of how AMD is extending it’s Direct Connect Architectureleadership and increasing density by connecting two CPUs on die.
August 13, 2004 Computation Products Group - High Density Data Center Committee 65
AMD Opteron™ ProcessorSingle Core vs. Dual Core Performance Estimates
FREQUENCY
PER
FOR
MA
NC
E
Single CoreDual Core
Note the rate of incline is steeper with dual core as each speed increase benefits 2 CPU’s!
Examples Built On an AMD64 Technology Foundation
August 13, 2004 Computation Products Group - High Density Data Center Committee 67
Rackable Cx000 Servers…
•High density with 176 processors per cabinet
•2U half-depth chassis enables back-to-back mounting for double density
•Front-facing I/O for easiest serviceability
•Choice of AC or DC power
•Supports fastest 64-bit processors
•Total component choice and flexibility
August 13, 2004 Computation Products Group - High Density Data Center Committee 68
Sun Microsystems V20z…
• One or two AMD Opteron 200 Series processors (models 242, 244, 248)
• Up to 16 GB of registered DDR1-333 SDRAM (4 DIMM slots per processor for a total of 8 DIMM slots; 128 bit plus ECC databus)
• Two 10/100/1000-Mbps BaseT Ethernet ports
• Two 64-bit PCI-X slots: – One full-length at 133 MHz – One half-length at 66 MHz
• One or two 36 GB, 73 GB, or 146 GB Ultra320 SCSI disk drives (Internal)
• Operating System: – Solaris 9 4/04 x86 Platform Edition (32-bit) – Red Hat Enterprise Linux 3 for AMD Opteron (32-
bit/64-bit) – SUSE Linux Enterprise Server 8 for AMD Opteron
(64-bit) – SUSE Linux 9 Professional (64-bit)– Microsoft Windows 2000– Microsoft Windows Server 2003
August 13, 2004 Computation Products Group - High Density Data Center Committee 69
HP ProLiant DL585…
• AMD Opteron™ Model 850 (2.4GHz/ 1MB), Model 848 (2.2GHz/ 1MB), Model 844 (1.8GHz/1MB), or Model 842 (1.6GHz/1MB)
• Up to 64GB of dual channel PC2100 DDR
• Smart Array 5i Plus controller, with 64MB memory and battery-backed write cache
• HP embedded NC7782 dual port PCI-X Gigabit Server Adapter
• 8 I/O slots: 6 x 64-bit/100 MHz PCI-X expansion slots and 2 x 64-bit/133MHz PCI-X slots
• HyperTransport (processor interconnect) runs at 800MHz with 6.4GB of bandwidth
• On-die memory controller which runs at processor core frequency
• 5.3 GB per processor of memory bandwidth
Best in class 4-way 32-bit performance supports the most complex, compute-intensive application environments
August 13, 2004 Computation Products Group - High Density Data Center Committee 70
IBM eServer 325 …
•1U rack-optimized chassis
•Up to 2 AMD OpteronTMprocessors
•1GB standard up to 12GB max PC2700 ECC DDR SDRAM memory
•Open-bay hot-swap Ultra320 SCSI or 40GB IDE drive standard; supports up to 2 hard disk drives
•Two 64-bit PCI-X slots
•Integrated dual Gigabit Ethernet ports
•Integrated systems management processor
August 13, 2004 Computation Products Group - High Density Data Center Committee 71
Cray XD1 … Exceptional Performance
Chassis Each Rack
Compute Processors 12 144
Performance 53 GFLOPS* 633 GFLOPS*
Aggregate Switching Capacity 96 GB/s 1152 GB/s
Interprocessor Latency 1.6 µsec 1.8 µsec
Aggregate Memory Bandwidth 77 GB/s 922 GB/s
Maximum Memory 96 GB 1.2 TB
Maximum Disk Storage 2 TB 24 TB
* Peak theoretical performance with 2.2 GHz AMD Opteron
August 13, 2004 Computation Products Group - High Density Data Center Committee 72
AMD Opteron™ ProcessorThermally Tested Server Solutions
Some Thermally Tested Server Barebones Solutions:
AW171* (Pedestal 1P)•BBATOA100-A01(1U/2P)•BBATOC100-A01(1U/2P Half Depth)•BBATOM208S-A01 (2U/2P)•BBATOM510S-A01 (5U/2P)
•Tyan S2880(Pedestal/2P)•Arima HDAMA(Pedestal/2P)•MSI 9131(Pedestal/2P)R3122 (1U)
•Transport GX28 (B2882) •TS-1881 (1U/2P)
K1-1000 (1U/2P)•Clusteron D (1U)•RC-0103 (1U 1+1Power Supplies)•RC-0104 (1U)•RC-0222 (2U)
Appro 1100H (1U/2P)
R4280A (2U/4P)•2100 (1U/2P)•4300 (3U/4P)
*Available only in EuropeGreen font = passed AMD Thermal Test
August 13, 2004 Computation Products Group - High Density Data Center Committee 73
Validated Server Program
A8440
•Designed by AMD manufactured by Celestica
•Available through the channel or direct (depending on volume). Current distributors in both NA and Europe.
•Validated for AMD Opteron X50 series
•Support for 256MB up to 2GB memory modules
US Support/Sales Phone: 1-866-350-8475Support Email: [email protected]: +44 1952 230 415
By email: [email protected] AMD Contact: Jon Green ([email protected])By phone: (512) 658-1498
A8440
A2210
August 13, 2004 Computation Products Group - High Density Data Center Committee 74
Infrastructure Information
• We encourage our customers & partners to visit AMD’s public website for infrastructure items including:
–Recommended Motherboards –Thermal Solution Guidelines –Thermally Tested Server Solutions –Memory Guidelines –Power Supply Guidelines –Tower Chassis Guidelines –Support Components
http://www.amd.com/configuration
-Online Processor Quick Reference Guidehttp://www.amd.com/processorquickrefguide
Q&AA test that you were all listening!
August 13, 2004 Computation Products Group - High Density Data Center Committee 76
Thank you!
August 13, 2004 Computation Products Group - High Density Data Center Committee 77