PCI Express®: Enabling New Opportunities for Graphics

Barry WagnerDirector of Technical MarketingNVIDIA Corporation

Session OutlineSession Outline

Increased bandwidth of PCI Express® enables a new direction for GPU design

Scalability of PCI Express creates a new category of gaming machines

PCI Express transition creates an opportunity for a new open standard for laptop graphics upgrades

PCI Express – Raising the Bar on Mainstream GraphicsPCI Express – Raising the Bar on Mainstream Graphics

Graphics companies utilize PCI Express bandwidth in place of dedicated graphics memory

Substantial cost saving for consumers

Graphics pipeline changed to allow rendering directly to system memory

Combination of HW and SW used to manage surface allocation

Users get better compatibility, stability, reliability from new, frequently improved drivers and GPU architectures at a lower price.

Rendering to System MemoryRendering to System Memory

Fundamental new direction in discrete GPU design

Direct rendering to system memory significantly reduces local graphics memory requirements

Better laptop performance at a lower power and cost due to fewer memories

Delivers latest graphics architectures and performance to affordable price points

Enabled by PCI Express

PCI Express Bandwidth ImprovementsPCI Express Bandwidth Improvements

Improved Write Bandwidth is Critical for Rendering Directly to System Memory

Bandwidth (GB/s)

READ1 WRITE2

AGP-8X 2 0.2PCI Express x16 4 4

1 READ from system memory to GPU2 WRITE from GPU to system memory

Benefits of PCI Express System Cache vs. Benefits of PCI Express System Cache vs. Conventional AGP ArchitectureConventional AGP Architecture

CoreLogicCoreLogic

SystemDRAM

SystemDRAM

3.2 GB/S

128MBGraphics Memory

4GB/scannot render efficiently to system memory

TraditionalGPU

16Mx16

16Mx16

CoreLogicCoreLogic

PCI ExpressGraphicsDevice

PCI ExpressGraphicsDevice

SystemDRAM

SystemDRAM

96MB dynamically

allocated for graphics

128MBGraphics Memory

13.6GB/s Peak bandwidth

32MB

4Mx324Mx32

4GB/s 4GB/s

5.6 GB/S

4GB/s Bi-directional PCI Express data paths enable efficient rendering to system memory

Typical 3D PipelineTypical 3D Pipeline

Triangle SetupTriangle Setup

L2 TexL2 Tex

Shader Instruction DispatchShader Instruction Dispatch

Fragment CrossbarFragment Crossbar

Graphics Memory

Graphics Memory

Graphics Memory

Graphics Memory

Z-CullZ-Cull

Shared Memory Through MMUShared Memory Through MMU

Render directly to system memory at full speed

Texture from system memory at full speed

Dynamically allocate surfaces anywhere

Present NVIDIA TurboCacheTM architecture

Triangle SetupTriangle Setup

L2 TexL2 Tex

Shader Instruction DispatchShader Instruction Dispatch

Fragment CrossbarFragment Crossbar

Graphics Memory

Graphics Memory

Graphics Memory

Graphics Memory

Z-CullZ-Cull

MMUMMU

SystemMemorySystemMemory

0

1

2

3

4

5

Shared Memory Enables Higher Performance at Shared Memory Enables Higher Performance at Lower PriceLower Price

Device QTY DRAM SpeedDRAM1

Price

Local Memory

CostShared Memory 4Mx32 DDR 2 350MHz 3.50$ 7.00$ Traditional 16Mx16 DDR 4 200 MHz 3.50$ 14.00$ Shared Memory 8Mx16 DDR 4 200MHz 2.00$ 8.00$ Traditional 16Mx16 DDR 8 200 MHz 3.50$ 28.00$

128MB

256MB

128MB Shared Memory

256MB Shared Memory

128MB

64-bit

Local Memory

256MB

128-bit

Local Memory

Integrated Graphics

1DRAM pricing changes constantly. A variety of market sources are available to confirm the approximate pricing reflected here. http://www.dramexchange.com is one such source.

3DMark05 10x7 0x AA / 0x Aniso

0

0.2

0.4

0.6

0.8

1

Integrated Graphics 128MB Traditional 64-bit GPU 128MB Shared Memory

WB 99 High-End

Disk

SysMark2004

PCMark 2004

Business Winstone

2004

Biz WS CC

WB 99 Business

Disk

Shared System Memory Design Outperforms Integrated Graphics and Traditional GPU architectures in a Variety of System Benchmarks

No Negative Impact to System PerformanceNo Negative Impact to System Performance

More Usable PCI Express Bandwidth = More Usable PCI Express Bandwidth = Faster PerformanceFaster Performance

0

0.2

0.4

0.6

0.8

1

1.2

1.4

Early PCI Express Core Logic New PCI Express Core Logic

4 GB/s

6 GB/s

3DMark05 10x7 0x AA / 0x Aniso

No need to wait for new applications to use the new bandwidth like we experienced with AGP

Key PointsKey Points

PCI Express has enabled a new direction in GPU Design

Graphics performance scales with usable PCI Express bandwidth improvements

Cost reductions for graphics memory enable more consumers to experience games the way they were meant to be played.

Laptops get better performance, lower costs, and longer battery life with system memory cache

PCI Express Returns Scalable Graphics PCI Express Returns Scalable Graphics Performance to the PCPerformance to the PC

AGP was NOT very scalableArchitected for a single graphics slot

Scaled frequency over the bus life-cycle

PCI Express planned for scalability Scalable bus widths : x1, x2, x4, x8, x16

Port Splitting

Down-shifting

Frequency scaling: 2.5GHz (Gen1), 5GHz (Gen2)

General purpose bus means multiple slots available for scaling performance

Port-Splitting in the Core LogicPort-Splitting in the Core Logic

Dividing a slot into multiple independent links

BenefitsAllows 2 or more GPU on a single x16 slot without the need for a bridge chip

Does not require multiple x16 slots on the motherboardEnables performance for wider install base of motherboards

DrawbacksHigher power density on a single card

May require larger than standard form factor cards

Optional feature in PCI Express specification so risk that a specific card may not work in every system.

Down-Shifting on the MotherboardDown-Shifting on the Motherboard

Wiring fewer lanes to a slot than the max x8 slots wired for x4 link width

x16 slots wired for x8 link widths

Why?There will always be a limited number of lanes

BenefitsCheapest way to get 2 slots that are capable of fitting the x16 edge connector used on graphics cards

2x the GPU performance with standard cards

Avoid inventory management of both x8 and x16 cards

DrawbackMotherboard has to support slot power level

Scalable Graphics ImpactScalable Graphics Impact

PCI Express High-End Graphics Specification75W from x16 slot

75W from HE power connector

Scalable solutions demand more power for the graphics subsystem

Graphics companies have already started to request the PCI-SIG® begin addressing a demand for greater than 150W cards.

When do Systems Benefit from Dual When do Systems Benefit from Dual GPU Designs? GPU Designs?

When you play Modern GPU intensive games at:

High Resolutions(1600x1200) with:Anti-Aliasing (4xAA)

Anisotropic Filtering (8xAF)

Example Applications that Scale: Battlefield Vietnam, Doom3, Far Cry

Game images are the property of their respective owners. All Rights Reserved.

Nvidia’s Design With Scalable PCI ExpressNvidia’s Design With Scalable PCI Express

GPU-to-GPU Interconnect1GB/s digital link

Provides pixel and synchronization information

Multiple system implementations supported

Rigid PCB with standard edge connectors

Flexible cables with standard edge connectors

Embedded designs

Optimized within SoftwareCan be enabled or disabled by driver and/or user

User may prefer to enable 4 displays for some applications in non-scaled mode

Solutions Possible for Any Solutions Possible for Any Compatible SystemCompatible System

Dual x16 slots fully-wired motherboard:2 compatible standard graphics cards with scalability support

Rigid or flexible cable between the cards

Single x16 slot without Port Splitting:2 GPUs on a single card plus a x16 to x8 PCI Express bridge

Inter-connect embedded in PCB

Single x16 slot with Port Splitting:2 GPUs on a single card

Inter-connect embedded in PCB

Dual x16 slots each down-shifted to x8:2 compatible graphics cards with scalability support

Rigid or flexible cable between the cards

Nvidia’s Dynamic Load Balancing TechnologyNvidia’s Dynamic Load Balancing Technology

HW and Driver work together to determine best algorithms to share workload between cards

Alternate Frame Rendering (AFR)

Split Frame Rendering (SFR)

Performance Improvement depends on ability to share work across cards

Works with almost any 3D application

Today some applications benefit more than othersSome popular applications show 1.7x – 2x increase

Future games will trend toward 2x increase as indicated by 3dMark05

NVIDIA SLINVIDIA SLITMTM

A new class of gaming performanceA new class of gaming performance

Benchmarks run at 1600x1200 4x/8x on nForce 4 SLI Motherboard with AMD Athlon 64FX

0%20%40%60%80%

100%120%140%160%180%200%

3DMark05 Doom 3 FlatOut Half-Life 2 Thief: DeadlyShadows

Warhammer40,000: DoW

SINGLE GPU DUAL GPU

Key Points:Key Points:

PCI Express provides for better scalability

Factor in dual GPU architectures into your graphic power budgets

A New Opportunity to Address Mobile Graphics Form A New Opportunity to Address Mobile Graphics Form FactorFactor

Desktop PC standard form factors has enabled growth and innovation

Rapid time to market

Range of choice addresses all market segments

Drives competition

The Mobile PC Platform has not benefited from a common form factor

Mobile PC development cycle is long

Per platform custom graphics integration is expensive

Custom design limits OEM/ODM choices & ultimately the consumer’s

Notebook Graphics Modules TodayNotebook Graphics Modules Today

Engineering resources are not leveraged

Custom BIOS

Custom connectors

Custom power delivery

Custom cooling

Custom Power Management

PCI Express is a catalyst for changePCI Express is a catalyst for change

PCI Express electrical changes require a new look at the platform

Robust differential links can tolerate modular design

No longer able to just send out the old footprint compatible laptop.

PCI Express brings new power management options

Scalable link widthsPower down links you don’t need

Notebook Platform has matured substantiallyPlatform requirements are better understood

Lessons learned from current module experience

What issues must be addressed by a new Mobile PC What issues must be addressed by a new Mobile PC Graphics Architecture?Graphics Architecture?

Broad industry participation

Open Architecture

A common host interface (PCI Express)

Enable broad range of power, thermal, & mechanical boundaries

Cover a variety of display output technologiesVGA, DVI, LVDS, Video (HDCP)

A common software layer and partitioning for the video BIOS and system BIOS

MXM – Mobile PCI Express ModuleMXM – Mobile PCI Express Module

Open architecture co-developed with many leading ODM/OEM

AOpen, Arima, Asustek, Clevo, FIC, Mitac, Quanta, Tatung, Uniwill, Wistron and more.

Enables a consistent graphics interface across all PCI Express notebooks

Supports up to x16 PCI Express

One design, many notebooks

Use different graphics solutions, from ANY vendor

Potential for consumer upgradeable graphics

Modules already designed for various NVIDIA, ATI, and S3 products

Key Points:Key Points:

The PCI Express transition is creating new opportunities for competition in the notebook PC

The notebook platform is evolving toward a common modular graphics form factor in many segments

MXM is one open architecture developed with the notebook industry to leverage the limited engineering resources.

Now is the time for the industry to come together on a common form factor.

Call To ActionCall To Action

Consider system memory cache solutions for your mainstream desktop and mobile markets

For maximum graphics performance, build systems with multiple x16 slots to enable best scalability of graphics hardware

Consider whether MXM is right for your next laptop design or purchasing decision

Let your PCI-SIG reps know you want the graphics industry power and scalability issues to continue to be addressed by the PCI-SIG.

For More InformationFor More Information

PCI Express Specificationswww.pcisig.com

NVIDIA PCI Express Product Informationhttp://www.nvidia.com/page/pci_express.html

NVIDIA TurboCache Technologyhttp://www.nvidia.com/page/turbocache.html

NVIDIA SLI Technologyhttp://www.nzone.com/object/nzone_sli_home.html

MXM Technologyhttp://www.nvidia.com/page/mxm.html

Community ResourcesCommunity Resources

Windows Hardware & Driver Central (WHDC)www.microsoft.com/whdc/default.mspx

Technical Communitieswww.microsoft.com/communities/products/default.mspx

Non-Microsoft Community Siteswww.microsoft.com/communities/related/default.mspx

Microsoft Public Newsgroupswww.microsoft.com/communities/newsgroups

Technical Chats and Webcastswww.microsoft.com/communities/chats/default.mspx

www.microsoft.com/webcasts

Microsoft Blogswww.microsoft.com/communities/blogs