Technical Brief

NVIDIA GeForce 8800 GPU

Architecture Overview

World’s First Unified DirectX 10 GPU

Delivering Unparalleled Performance and
Image Quality
November 2006
TB-02787-001_v0.9

TB-02787-001_v01 i
NVIDIA GeForce 8800 Architecture Technical Brief

ii TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

Table of Contents

Preface.......................................................................................................................... vii
GeForce 8800 Architecture Overview ............................................................................1
Unified, Massively Parallel Shader Design............................................................................ 1
DirectX 10 Native Design................................................................................................... 3
Lumenex Engine: Industry-Leading Image Quality............................................................... 5
SLI Technology ................................................................................................................ 7
Quantum Effects GPU-Based Physics .................................................................................. 7
PureVideo and PureVideo HD............................................................................................. 9
Extreme High Definition Gaming (XHD) ............................................................................ 11
Built for Microsoft Windows Vista ..................................................................................... 12
CUDA: Compute Unified Device Architecture.................................................................. 12
The Four Pillars........................................................................................................... 15
The Classic GPU Pipeline… A Retrospective.................................................................17
GeForce 8800 Architecture in Detail ............................................................................19
Unified Pipeline and Shader Design .................................................................................. 20
Unified Shaders In-Depth ............................................................................................ 21
Stream Processing Architecture........................................................................................ 25
Scalar Processor Design Improves GPU Efficiency .......................................................... 27
Lumenex Engine: High-Quality Antialiasing, HDR, and Anisotropic Filtering .......................... 27
Decoupled Shader/Math, Branching, and Early-Z ............................................................... 31
Decoupled Shader Math and Texture Operations ........................................................... 31
Branching Efficiency Improvements .............................................................................. 32
Early-Z Comparison Checking....................................................................................... 33
GeForce 8800 GTX GPU Design and Performance........................................................35
Host Interface and Stream Processors .............................................................................. 36
Raw Processing and Texturing Filtering Power............................................................... 36
ROP and Memory Subsystems...................................................................................... 37
Balanced Architecture ..................................................................................................... 38

TB-02787-001_v1.0 iii
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

DirectX 10 Pipeline .......................................................................................................39

Virtualization and Shader Model 4 .................................................................................... 39
Stream Output ............................................................................................................... 41
Geometry Shaders .......................................................................................................... 42
Improved Instancing....................................................................................................... 43
Vertex Texturing ......................................................................................................... 44
The Hair Challenge ..................................................................................................... 44
Conclusion ....................................................................................................................45

iv TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

List of Figures

Figure 1. GeForce 8800 GTX Block Diagram........................................................................ 2

Figure 2. DirectX 10 game “Crysis” with both HDR lighting and antialiasing ........................... 4
Figure 3. NVIDIA Lumenex engine delivers incredible realism............................................... 6
Figure 4. NVIDIA SLI technology ....................................................................................... 7
Figure 5. Quantum Effects ................................................................................................ 8
Figure 6. HQV benchmark results for GeForce 8800 GPUs .................................................. 10
Figure 7. PureVideo vs. the competition ........................................................................... 10
Figure 8. Extreme High Definition widescreen gaming ....................................................... 11
Figure 9. CUDA thread computing pipeline........................................................................ 13
Figure 10. CUDA thread computing parallel data cache........................................................ 14
Figure 11. Classic GPU pipeline.......................................................................................... 17
Figure 12. GeForce 8800 GTX block diagram ...................................................................... 20
Figure 13. Classic vs. Unified Shader Architecture ............................................................... 21
Figure 14. Characteristic pixel and vertex shader workload variation over time ...................... 22
Figure 15. Fixed shader performance characteristics ........................................................... 23
Figure 16. Unified shader performance characteristics ......................................................... 24
Figure 17. Conceptual unified shader execution framework.................................................. 25
Figure 18. Streaming processors and texture units .............................................................. 26
Figure 19. Coverage sampling antialiasing (4× MSAA vs. 16× CSAA) .................................... 28
Figure 20. Isotropic trilinear mipmapping (left) vs. anisotropic trilinear mimapping (right) ...... 29
Figure 21. Anisotropic filtering comparison (GeForce 7 Series on left, and GeForce 8 Series or
right using default anisotropic Texture Filtering).................................................. 30
Figure 22. Decoupling texture and math operations ............................................................ 31
Figure 23. GeForce 8800 GPU pixel shader branching efficiency ........................................... 32
Figure 24. Example of Z-buffering ..................................................................................... 33
Figure 25. Example of early-Z technology........................................................................... 34
Figure 26. GeForce 8800 GTX block diagram ...................................................................... 35
Figure 27. Texture fill performance of GeForce 8800 GTX .................................................... 37
Figure 28. Direct3D 10 pipeline ......................................................................................... 41
Figure 29. Instancing at work—numerous characters rendered ............................................ 43

TB-02787-001_v1.0 v
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

List of Tables

Table 1. Shader Model Progression ................................................................................. 40

Table 2. Hair Algorithm Comparison of DirectX 9 and DirectX 10 ....................................... 44

vi TB-02787-001_v1.0
November 8, 2006
 Preface

Welcome to our technical brief describing the NVIDIA® GeForce® 8800 GPU
architecture.
We have structured the material so that the initial few pages discuss key GeForce
8800 architectural features, present important DirectX 10 capabilities, and describe
how GeForce 8 Series GPUs and DirectX 10 work together. If you read no further,
you will have a basic understanding of how GeForce 8800 GPUs enable
dramatically enhanced 3D game features, performance, and visual realism.
In the next section we go much deeper, beginning with operations of the classic
GPU pipeline, followed by showing how GeForce 8800 GPU architecture radically
changes the way GPU pipelines operate. We describe important new design features
of GeForce 8800 architecture as it applies to both the GeForce 8800 GTX and the
GeForce 8800 GTS GPUs. Throughout the document, all specific GPU design and
performance characteristics are related to the GeForce 8800 GTX.
Next we’ll look a little closer at the new DirectX 10 pipeline, including a
presentation of key DirectX 10 features and Shader Model 4.0. Refer to the
NVIDIA technical brief titled Microsoft DirectX 10: The Next-Generation Graphics API
(TP-02820-001) for a detailed discussion of DirectX 10 features.
We hope you find this information informative.

TB-02787-001_v1.0 vii
November 4, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

viii TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture
Overview

Based on the revolutionary new NVIDIA® GeForce® 8800 architecture, NVIDIA’s

powerful GeForce 8800 GTX graphics processing unit (GPU) is the industry’s first
fully unified architecture-based DirectX 10–compatible GPU that delivers incredible
3D graphics performance and image quality. Gamers will experience amazing
Extreme High Definition (XHD) game performance with quality settings turned to
maximum, especially with NVIDIA SLI® configurations using high-end NVIDIA
nForce® 600i SLI motherboards.

Unified, Massively Parallel

Shader Design
The GeForce 8800 GTX GPU implements a massively parallel, unified shader
design consisting of 128 individual stream processors running at 1.35 GHz. Each
processor is capable of being dynamically allocated to vertex, pixel, geometry, or
physics operations for the utmost efficiency in GPU resource allocation and
maximum flexibility in load balancing shader programs. Efficient power utilization
and management delivers industry-leading performance per watt and performance
per square millimeter.

TB-02787-001_v1.0 1
November 4, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Figure 1. GeForce 8800 GTX Block Diagram

Don’t worry—we’ll describe all the gory details of Figure 1 very shortly! Compared
to the GeForce 7900 GTX, a single GeForce 8800 GTX GPU delivers 2× the
performance on current applications, with up to 11× scaling measured in certain
shader operations. As future games become more shader intensive, we expect the
GeForce 8800 GTX to surpass DirectX 9–compatible GPU architectures in
performance.
In general, shader-intensive and high dynamic-range (HDR)–intensive applications
shine on GeForce 8800 architecture GPUs. Teraflops of raw floating-point
processing power are combined to deliver unmatched gaming performance, graphics
realism, and real-time, film-quality effects.
The groundbreaking NVIDIA® GigaThread™ technology implemented in GeForce
8 Series GPUs supports thousands of independent, simultaneously executing
threads, maximizing GPU utilization.

2 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

The GeForce 8800 GPU’s unified shader architecture is built for extreme 3D
graphics performance, industry-leading image quality, and full compatibility with
DirectX 10. Not only do GeForce 8800 GPUs provide amazing DirectX 10 gaming
experiences, but they also deliver the fastest and best quality DirectX 9 and
OpenGL gaming experience today. (Note that Microsoft Windows Vista is required
to utilize DirectX 10).
We’ll briefly discuss DirectX 10 features supported by all GeForce 8800 GPUs, and
then take a look at important new image quality enhancements built into every
GeForce 8800 GPU. After describing other essential GeForce 8800 Series
capabilities, we’ll take a deep dive into the GeForce 8800 GPU architecture,
followed by a closer look at the DirectX 10 pipeline and its features.

DirectX 10 Native Design

DirectX 10 represents the most significant step forward in 3D graphics APIs since
the birth of programmable shaders. Completely built from the ground up, DirectX
10 features powerful geometry shaders, a new “Shader Model 4” programming
model with substantially increased resources and improved performance, a highly
optimized runtime, texture arrays, and numerous other features that unlock a whole
new world of graphical effects (See “DirectX 10 Pipeline” later in this document).
GeForce 8 Series GPUs include all the required hardware functionality defined in
Microsoft’s Direct3D 10 (DirectX 10) specification and full support for the DirectX
10 unified shader instruction set and Shader Model 4 capabilities. The GeForce
8800 GTX is not only the first shipping DirectX 10 GPU, but it was also the
reference GPU for DirectX 10 API development and certification. (For more details
on DirectX 10, refer to Microsoft DirectX 10: The Next-Generation Graphics API.)
New features implemented in GeForce 8800 Series GPUs that work in concert with
DirectX 10 features include geometry shader processing, stream output, improved
instancing, and support for the DirectX 10 unified instruction set. GeForce 8 Series
GPUs and DirectX 10 also provide the ability to reduce CPU overhead, shifting
more graphics rendering load to the GPU.

TB-02787-001_v1.0 3
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Courtesy of Crytek

Figure 2. DirectX 10 game “Crysis” with both HDR lighting and

antialiasing
DirectX 10 games running on GeForce 8800 GPUs deliver rich, realistic scenes;
increased character detail; and more objects, vegetation, and shadow effects in
addition to natural silhouettes and lifelike animations.
PC-based 3D graphics is raised to the next level with GeForce 8800 GPUs
accelerating DirectX 10 games.

4 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

Lumenex Engine: Industry-

Leading Image Quality
Image quality is significantly improved on GeForce 8800 GPUs with the NVIDIA
Lumenex™ engine. Advanced new antialiasing technology provides up to 16× full-
screen multisampled antialiasing quality at near 4× multisampled antialiasing
performance using a single GPU.
High dynamic-range (HDR) lighting capability in all GeForce 8800 Series GPUs
supports 128-bit precision (32-bit floating-point values per component), permitting
true-to-life lighting and shadows. Dark objects can appear very dark—and bright
objects can be very bright—with visible details present at both extremes, in addition
to completely smooth gradients rendered in between.
HDR lighting effects can be used in concert with multisampled antialiasing on
GeForce 8 Series GPUs. Plus, the addition of angle-independent anisotropic
filtering, combined with considerable HDR shading horsepower, provides
outstanding image quality. In fact, antialiasing can be used in conjunction with both
FP16 (64-bit color) and FP32 (128-bit color) render targets.
The following image of model Adrianne Curry was rendered using a GeForce 8800
GTX GPU, and clearly illustrates the realistic effects made possible by the NVIDIA
Lumenex engine.

TB-02787-001_v1.0 5
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

(Image of model Adrianne Curry rendered on a GeForce 8800 GTX GPU)

Figure 3. NVIDIA Lumenex engine delivers incredible realism

An entirely new 10-bit display architecture works in concert with 10-bit DACs to
deliver over a billion colors (compared to 16.7 million in the prior generation),
permitting incredibly rich and vibrant photos and videos. With the next generation
of 10-bit content and displays, the Lumenex engine will be able to display images of
amazing depth and richness.
For more details on GeForce 8800 GPU image quality improvements, refer to
Lumenex Engine: The New Standard in GPU Image Quality (TB-02824-001).

6 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

SLI Technology
NVIDIA’s SLI technology is the industry’s leading multi-GPU technology. It
delivers up to 2× the performance of a single GPU configuration for unequaled
gaming experiences by allowing two graphics cards to run in parallel on a single
motherboard. The must-have feature for performance PCI Express graphics, SLI
dramatically scales performance on today’s hottest games. Running two GeForce
8800 GTX boards in an SLI configuration allows extremely high image-quality
settings at extreme resolutions.

Figure 4. NVIDIA SLI technology

Quantum Effects GPU-Based

Physics
NVIDIA Quantum Effects™ technology enables more physics effects to be
simulated and rendered on the GPU. Specifically, GeForce 8800 GPU stream
processors excel at physics computations, and up to 128 processors deliver a
staggering floating-point computational ability that results in amazing performance
and visual effects. Games can implement much more realistic smoke, fire, and
explosions. Also, lifelike movement of hair, fur, and water can be completely
simulated and rendered by the graphics processor. The CPU is freed up to run the
game engine and artificial intelligence (AI), thus improving overall gameplay. Expect
to see far more physics simulations in DirectX 10 games running on GeForce 8800
GPUs.

TB-02787-001_v1.0 7
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Figure 5. Quantum Effects

8 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

PureVideo and PureVideo HD

The NVIDIA PureVideo™ HD capability is built into every GeForce 8800 Series
GPU and enables the ultimate HD DVD and Blu-ray viewing experience with
superb picture quality, ultra-smooth movie playback, and low CPU utilization. High-
precision subpixel processing enables videos to be scaled with great precision,
allowing low-resolution videos to be accurately mapped to high-resolution displays.
PureVideo HD is comprised of dedicated GPU-based video processing hardware
(SIMD vector processor, motion estimation engine, and HD video decoder),
software drivers, and software-based players that accelerate decoding and enhance
image quality of high-definition video in H.264, VC-1, WMV/WMV-HD, and
MPEG-2 HD formats.
PureVideo HD can deliver 720p, 1080i, and 1080p high-definition output and
support for both 3:2 and 2:2 pull-down (inverse telecine) of HD interlaced content.
PureVideo HD on GeForce 8800 GPUs now provides HD noise reduction and HD
edge enhancement.
PureVideo HD adjusts to any display and uses advanced techniques (found only on
high-end consumer players and TVs) to make standard and high-definition video
look crisp, smooth, and vibrant, regardless of whether videos are watched on an
LCD, plasma, or other progressive display type.
AACS-protected Blu-ray or HD DVD movies can be played on systems with
GeForce 8800 GPUs using AACS-compliant movie players from CyberLink,
InterVideo, and Nero that utilize GeForce 8800 GPU PureVideo features.
All GeForce 8800 GPUs are HDCP-capable, meeting the security specifications of
the Blu-ray Disc and HD DVD formats and allowing the playback of encrypted
movie content on PCs when connected to HDCP-compliant displays.
GeForce 8800 Series GPUs also readily handle standard definition PureVideo
formats such as WMV and MPEG-2 for high-quality playback of computer-
generated video content and standard DVDs. In the popular industry-standard
HQV Benchmark (www.hqv.com), which evaluates standard definition video de-
interlacing, motion correction, noise reduction, film cadence detection, and detail
enhancement, all GeForce 8800 GPUs achieve an unrivaled 128 points out of 130 points!

TB-02787-001_v1.0 9
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Figure 6. HQV benchmark results for GeForce 8800 GPUs

PureVideo and PureVideo HD are programmable technologies that can adapt to

new video formats as they are developed, providing a future-proof video solution.

Figure 7. PureVideo vs. the competition

GeForce 8800 GPUs support various TV-out interfaces such as composite,

S-video, component, and DVI. HD resolutions up to 1080p are supported
depending on connection type and TV capability.

10 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

Extreme High Definition

Gaming (XHD)
All GeForce 8800–based GPUs are designed for Extreme High Definition gaming
(XHD), where games can be played at high widescreen resolutions up to
2560×1600. XHD has over seven times the picture clarity of native 1080i HD
televisions and double the picture clarity of the 1080p HD format. XHD widescreen
resolution allows users to see more of their PC games, enhance their video editing,
and even add useful extra screen real estate to applications such as Google Earth.
The dual-link DVI outputs on GeForce 8800 GTX boards enable XHD gaming up
to 2560×1600 resolution with very playable frame rates. SLI configurations allow
dialing up eye-candy to new levels of details never seen in the past, all with playable
frame rates.

Figure 8. Extreme High Definition widescreen gaming

TB-02787-001_v1.0 11
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Built for Microsoft Windows

Vista
GeForce 8800 GPU architecture is actually NVIDIA’s fourth-generation GPU
architecture built for Microsoft® Windows Vista™ technology, and gives users the
best possible experience with the Windows Aero 3D graphical user interface and full
DirectX 10 hardware support. GeForce 8800 GPUs support for Vista includes
Windows Display Driver Model (WDDM), Vista’s Desktop Windows Manager
(DWM) composited desktop, the AERO interface using DX9 3D graphics, fast
context switching, GPU resource virtualization support, and OpenGL Installable
Client Driver (ICD) support (both older XP ICDs and newer Vista ICDs).

CUDA: Compute Unified Device Architecture

All GeForce 8800 GPUs include the revolutionary new NVIDIA CUDA™ built-in
technology, which provides a unified hardware and software solution for data-
intensive computing. Key highlights of CUDA technology are as follows:
‰ New “Thread Computing” processing model that takes advantage of massively
threaded GeForce 8800 GPU architecture, delivering unmatched performance
for data-intensive computations.
‰ Computing threads that can communicate and cooperate on the GPU.
‰ Standard C language interface for a simplified platform for complex
computational problems.
‰ Architecture that complements traditional CPUs by providing additional
processing capability for inherently parallel applications.
‰ Use of GPU resources in a different manner than graphics processing as seen in
Figure 9, but both CUDA threads and graphics threads can run on the GPU
concurrently if desired.

12 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

Figure 9. CUDA thread computing pipeline

CUDA enables new applications with a standard platform for extracting valuable
information from vast quantities of raw data, and provides the following key
benefits in this area:
‰ Enables high-density computing to be deployed on standard enterprise
workstations and server environments for data-intensive applications.
‰ Divides complex computing tasks into smaller elements that are processed
simultaneously in the GPU to enable real-time decision making.
‰ Provides a standard platform based on industry-leading NVIDIA hardware and
software for a wide range of high data bandwidth, computationally intensive
applications.
‰ Combines with multicore CPU systems to provide a flexible computing
platform.
‰ Controls complex programs and coordinates inherently parallel computation on
the GPU processed by thousands of computing threads.

CUDA’s high-performance, scalable computing architecture solves complex parallel

problems 100× faster than traditional CPU-based architectures:
‰ Up to of 128 parallel 1.35 GHz compute cores in GeForce 8800 GTX GPUs
harness massive floating-point processing power, enabling maximum
application performance.
‰ Thread computing scales across NVIDIA’s complete line of next-generation
GPUs—from embedded GPUs to high-performance GPUs that support
hundreds of processors.

TB-02787-001_v1.0 13
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

‰ NVIDIA SLI™ technology allows multiple GPUs to distribute computing to

provide unparalleled compute density.
‰ Enables thread computing to be deployed in any industry-standard
environment.
‰ Parallel Data Cache stores information on the GPU so threads can share data
entirely within the GPU for dramatically increased performance and flexibility.

Figure 10. CUDA thread computing parallel data cache

‰ Thread Execution Manager efficiently schedules and coordinates the execution

of thousands of computing threads for precise computational execution.

CUDA SDK unlocks the power of the GPU using industry-standard C language:
‰ Industry-standard C compiler simplifies software for complex computational
problems.
‰ Complete development solution includes an industry-standard C compiler,
standard math libraries, and a dedicated driver for thread computing on either
Linux or Windows.
‰ Full support of hardware debugging and a profiler for program optimization.
‰ NVIDIA “assembly for computing” (NVasc) provides lower-level access to the
GPU for computer language development and research applications.

14 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture Overview

The Four Pillars

Overall, the GeForce 8800 GPU Series is best defined by these four major
categories:
‰ Outstanding performance with a unified shader design
NVIDIA GigaThread technology and overall thread computing capability
delivers the absolute best GPU performance for 3D games.
‰ DirectX 10 compatibility
GeForce 8800 Series GPUs are the first shipping GPUs that support all
DirectX 10 features.
‰ Significantly improved image quality
NVIDA Lumenex engine technology provides top-quality antialiasing,
anisotropic filtering, and HDR capabilities enabling rich, lifelike detail in 3D
games.
‰ High-performance GPU physics and GPU computing capability
NVIDIA Quantum Effects permits billions of physics operations to be
performed on the GPU, enabling amazing new effects and providing massive
floating-point computing power for a variety of high-end calculation-intensive
applications.

That’s the high-level overview. Now it is time to go deep!

In the following sections, we first review classic GPU pipeline design and compare
it to the new unified pipeline and shader architecture used by GeForce 8800 GPUs.
We then discuss stream processors and scalar versus vector processor design, so
you’ll better understand GeForce 8800 GPU technology. Next, we’ll present a high-
level view of the GeForce 8800 GTX GPU architecture, followed by many of the
new features that apply to all GeForce 8800 GPUs. All the while we’ll provide
specific references to GeForce 8800 GTX design and performance characteristics.
The final section looks at important aspects of the DirectX 10 pipeline and
programming model, and how they relate to the GeForce 8800 GPU architecture.

TB-02787-001_v1.0 15
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

This page is blank.

16 TB-02787-001_v1.0
November 8, 2006
 The Classic GPU Pipeline…
A Retrospective

Before we examine architectural details of a GeForce 8800 GPU, we should explain

how the classic GPU internal pipeline operations have worked over the years.
When discussing GPU hardware architectures, dataflow, and pipeline operations, it’s
often good to start from the top, where data is fed from the CPU to the GPU, and
work our way down through multiple processing stages until a pixel is finally drawn
on the screen.
To date, GPUs have utilized traditional pipelined designs, as shown in Figure 11,
which broadly depicts the GPU pipeline evolution through the DX9 generation.

Figure 11. Classic GPU pipeline

TB-02787-001_v1.0 17
November 4, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

After the GPU receives vertex data from the host CPU, the vertex stage is the first
major stage. Back in the DirectX 7 timeframe, fixed-function transform and lighting
hardware operated at this stage (such as with NVIDIA’s GeForce 256 in 1999), and
then programmable vertex shaders came along with DirectX 8. This was followed
by programmable pixel shaders in DirectX 9 Shader Model 2, and dynamic flow
control in DirectX 9 Shader Model 3. DirectX 10 expands programmability features
much further, and shifts more graphics processing to the GPU, significantly
reducing CPU overhead.
The next step in the classic pipeline is the setup, where vertices are assembled into
primitives such as triangles, lines, or points. The primitives are then converted by
the rasterization stage into pixel fragments (or just “fragments”), but are not
considered full pixels at this stage. Fragments undergo many other operations such
as shading, Z-testing, possible frame buffer blending, and antialiasing. Fragments are
finally considered pixels when they have been written into the frame buffer.
As a point of confusion, the “pixel shader” stage should technically be called the
“fragment shader” stage, but we’ll stick with pixel shader as the more generally
accepted term. In the past, the fragments may have only been flat shaded or have
simple texture color values applied. Today, a GPU’s programmable pixel shading
capability permits numerous shading effects to be applied while working in concert
with complex multitexturing methods.
Specifically, shaded fragments (with color and Z values) from the pixel stage are
then sent to the ROP (Raster Operations in NVIDIA parlance). The ROP stage
corresponds to the “Output Merger” stage of the DirectX 10 pipeline, where Z-
buffer checking ensures only visible fragments are processed further, and visible
fragments, if partially transparent, are blended with existing frame buffer pixels and
antialiased. The final processed pixel is sent to the frame buffer memory for scanout
and display to the monitor.
The classic GPU pipeline has basically included the same fundamental stages for the
past 20 years, but with significant evolution over time. Many processing constraints
and limitations exist with classic pipeline architectures, as did variations in DirectX
implementations across GPUs from different vendors.
A few notable problems of pre-DirectX 10 classic pipelines include the following:
limited reuse of data generated within the pipeline to be used as input to a
subsequent processing step; high state change overhead; excessive variation in
hardware capabilities (requiring different application code paths for different
hardware); instruction set and data type limitations (such as lack of integer
instructions and weakly defined floating point precision); inability to write results to
memory in mid-pipeline and read them back into the top of the pipeline; and
resource limitations (registers, textures, instructions per shader, render targets, and
so on.) 1
Let’s proceed and see how the GeForce 8800 GPU architecture totally changes the
way data is processed in a GPU with it unified pipeline and shader architecture.

1 The Direct3D® 10 System, David Blythe, Microsoft Corporation.

18 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800
Architecture in Detail

When NVIDIA’s engineers started designing the GeForce 8800 GPU architecture
in the summer of 2002, they set forth a number of important design goals. The top
four goals were quite obvious:
‰ Significantly increase performance over current-generation GPUs.
‰ Notably improve image quality.
‰ Deliver powerful GPU physics and high-end floating-point computation ability.
‰ Provide new enhancements to the GPU pipeline (such as geometry shading and
stream output), while working collaboratively with Microsoft to define features
for the next major version of Direct X (DirectX 10 and Windows Vista).

In fact, many key GeForce 8800 architecture and implementation goals were
specified in order to make GeForce 8800–class GPUs most efficient for DirectX 10
applications, while also providing the highest performance for existing applications
using DirectX 9, OpenGL, and earlier DirectX versions.
The new GPU architecture would need to perform well on a variety of applications
using different mixes of pixel, vertex, and geometry shading in addition to large
amounts of high quality texturing.
The result was the GeForce 8800 GPU architecture that initially included two
specific GPUs—the high-end GeForce 8800 GTX and the slightly downscaled
GeForce 8800 GTS.
Figure 12 again presents the overall block diagram of the GeForce 8800 GTX for
readers who would like to see the big picture up front.
But fear not, we’ll start by describing the key elements of the GeForce 8800
architecture followed by looking at the GeForce 8800 GTX in more detail, where
we will again display this “most excellent” diagram and discuss some of its specific
features.

TB-02787-001_v1.0 19
November 4, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Figure 12. GeForce 8800 GTX block diagram

Unified Pipeline and Shader

Design
Recall the classic pipeline model with its dataflow starting at the top, where vertices
with various attributes, indices, commands, and textures are passed into the GPU
from the CPU. Major processing stages follow in a fairly linear manner, including
vertex shading, pixel shading, raster operations, and writing final pixels to the frame
buffer. In fact, the GeForce 7 Series GPUs had many physical pipeline stages per
major processing stage. Just the pixel shader stage of GeForce 7 GPUs had over 200
sequential pipeline stages!
With its unified pipeline and shader architecture, the GeForce 8800 GPU design
significantly reduces the number of pipeline stages and changes the sequential flow
to be more looping oriented. Inputs are fed to the top of the unified shader core,
and outputs are written to registers and then fed back into the top of the shader
core for the next operation.

20 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

In the generalized unified GPU diagram shown in Figure 13, the classic pipeline
uses discrete shader types represented in different colors, where data flows
sequentially down the pipeline through different shader types. The illustration on
the right depicts a unified shader core with one or more standardized, unified shader
processors.

Figure 13. Classic vs. Unified Shader Architecture

Data coming in the top left of the unified design (such as vertices), are dispatched to
the shader core for processing, and results are sent back to the top of the shader
core, where they are dispatched again, processed again, looped to the top, and so on
until all shader operations are performed and the pixel fragment is passed on to the
ROP subsystem.

Unified Shaders In-Depth

The GeForce 8800 design team realized that extreme amounts of hardware-based
shading horsepower would be necessary for high-end DirectX 10 3D games. While
DirectX 10 specifies a unified instruction set, it does not demand a unified GPU
shader design, but NVIDIA GeForce 8800 engineers believed a unified GPU shader
architecture made the most sense to allow effective DirectX 10 shader program
load-balancing, efficient GPU power utilization, and significant improvement to the
GPU architectural efficiency.
Note that the GeForce 8800 unified shaders can be also be used with DirectX 9,
OpenGL, and older DirectX versions. No restrictions or fixed numbers of unified
shading units need to be dedicated to pixel or vertex processing for any of the API
programming models.

TB-02787-001_v1.0 21
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

In general, numerous challenges had to be overcome with such a radical new design
over the four-year GeForce 8800 GPU development timeframe.
Looking more closely at graphics programming, we can safely say that in general the
number of pixels outnumbers vertices by a wide margin, which is why you saw a
much larger number of pixel shader units versus vertex shader units in prior fixed-
shader GPU architectures. But different applications do have different shader
processing requirements at any given point in time—some scenes may be pixel-
shader intensive and other scenes may be vertex-shader intensive. Figure 14 shows
the variations in vertex and pixel processing over time in a particular application.

Figure 14. Characteristic pixel and vertex shader workload

variation over time

In a GPU with a fixed number of specific types of shader units, restrictions are
placed on operating efficiency, attainable performance, and application design.
Figure 15 shows a theoretical GPU with a fixed number of 4 vertex shader units and
8 pixel shader units, or a total of 12 shader units altogether.

22 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

Figure 15. Fixed shader performance characteristics

In Figure 15, the top scenario shows a scene that is vertex shader-intensive, which
can only attain performance as fast as the maximum number of vertex units, which
in this case is “4.” In the bottom scenario, the scene is pixel shader-intensive, which
might be due to various complex lighting effects for the water. In this case, it is
pixel-shader limited and can only attain a maximum performance of “8,” equal to
the number of pixel shader units, which is the bottleneck in this case. Both
situations are not optimal because hardware is idle and performance is left on the
table, so to speak. Also, it’s not efficient from a power (performance/watt) or die
size and cost (performance/sqmm) perspective.
In Figure 16, with a unified shader architecture, at any given moment when an
application might be vertex-shader intensive, you can see the majority of unified
shader processors are applied to processing vertex data, and in this case, the overall
performance is increased to “11.” Similarly, if its pixel shader heavy, the majority of
unified shader units can be applied to pixel processing, also attaining a score of “11”
in our example.

TB-02787-001_v1.0 23
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Figure 16. Unified shader performance characteristics

Unified stream processors (SPs) in GeForce 8800 GPUs can process vertices, pixels,
geometry or physics—they are effectively general purpose floating-point processors.
Different workloads can be mapped to the processors, as shown in Figure 17.
Note that geometry shading is a new feature of the DirectX 10 specification that we
cover in detail later in the DirectX 10 section. The GeForce 8800 unified stream
processors can process geometry shader programs, permitting a powerful new range
of effects and features, while reducing dependence on the CPU for geometry
processing.
The GPU dispatch and control logic can dynamically assign vertex, geometry,
physics, or pixel operations to available SPs without worrying about fixed numbers
of specific types of shader units. In fact, this feature is just as important to
developers, who need not worry as much that certain aspects of their code might be
too pixel-shader intensive or too vertex-shader intensive.
Not only does a unified shader design assist in load-balancing shader workloads, but
it actually helps redefine how a graphics pipeline is organized. In the future, it is
possible that other types of workloads can be run on a unified stream processor.

24 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

Figure 17. Conceptual unified shader execution framework

Stream Processing Architecture

Key to the GeForce 8800 architecture is the use of numerous scalar stream
processors (SPs) to perform shader operations. Stream processors are highly
efficient computing engines that perform calculations on an input stream, while
producing an output stream that can be used by other stream processors. Stream
processors can be grouped in close proximity, and in large numbers, to provide
immense parallel processing power.
Generally, specialized high-speed instruction decode and execution logic is built into
a stream processor, and similar operations are performed on the different elements
of a data stream. On-chip memory is typically used to store output of a stream
processor, and the memory can be quickly read as input by other stream processors
for subsequent processing. SIMD (single instruction/multiple data) instructions can
be implemented across groupings of stream processors in an efficient manner, and
massively parallel stream processor clusters are well-suited for processing graphics
data streams.

TB-02787-001_v1.0 25
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

With the GeForce 8800 architecture, the image in Figure 18 shows a collection of
stream processors (SPs) with associated numbers of Texture Filtering (TF), Texture
Addressing (TA), and cache units to ensure a balanced design. The ratios of unit
types shown below in a subset slice of a typical GeForce 8800 GPU are maintained
when scaling up to 128 SPs specifically in a GeForce 8800 GTX GPU.

Figure 18. Streaming processors and texture units

Each GeForce 8800 GPU stream processor is fully generalized, fully decoupled,
scalar (see “Scalar Processor Design Improves GPU Efficiency”), can dual-issue a
MAD and a MUL, and supports IEEE 754 floating-point precision.
The stream processors are a critical component of NVIDIA GigaThread
technology, where thousands of threads can be in flight within a GeForce 8800
GPU at any given instant. GigaThread technology keeps SPs fully utilized by
scheduling and dispatching various types of threads (such as pixel, vertex, geometry,
and physics) for execution.
All stream processors are driven by a high-speed clock domain that is separate from
the core clock that drives the rest of the chip. For example, the GeForce 8800 GTX
core clock is 575 MHz and its stream processors run at 1.35 GHz delivering
exceptionally high shader performance.

26 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

Scalar Processor Design Improves GPU Efficiency

Leading GPUs to date have used vector processing units because many operations
in graphics occur with vector data (such as R-G-B-A components operating in pixel
shaders or 4×4 matrices for geometry transforms in vertex shaders). However, many
scalar operations also occur. During the early GeForce 8800 architecture design
phases, NVIDIA engineers analyzed hundreds of shader programs that showed an
increasing use of scalar computations. They realized that with a mix of vector and
scalar instructions, especially evident in longer, more complex shaders, it’s hard to
efficiently utilize all processing units at any given instant with a vector architecture.
Scalar computations are difficult to compile and schedule efficiently on a vector
pipeline.
NVIDIA and ATI vector-based GPUs have used shader hardware that permits dual
instruction issue. Recent ATI GPUs use a “3+1” design, allowing a single issue of a
four-element vector instruction, or a dual issue of a three-element vector instruction
plus a scalar instruction. NVIDIA GeForce 6x and GeForce 7x GPUs are more
efficient with 3+1 and 2+2 dual-issue design. But they are still not as efficient as a
GeForce 8800 GPU scalar design, which can issue scalar operations to its scalar
processors with 100 percent shader processor efficiency. NVIDIA engineers have
estimated as much as 2× performance improvement can be realized from a scalar
architecture that uses 128 scalar processors versus one that uses 32 four-component
vector processors, based on architectural efficiency of the scalar design. (Note that
vector-based shader program code is converted to scalar operations inside a
GeForce 8800 GPU to ensure complete efficiency.)

Lumenex Engine: High-Quality

Antialiasing, HDR, and Anisotropic
Filtering
NVIDIA’s Lumenex engine in GeForce 8800 GPUs implements entirely new and
very high-quality antialiasing (AA) and anisotropic filtering (AF) technologies. The
new antialiasing technology uses “coverage samples” and color and Z samples, and
is Coverage Sampling Antialiasing (CSAA).
CSAA technology provides much higher levels of antialiasing than possible with
present-day GPUs, with performance characteristics similar to 4× or 8×
multisampled AA. New CSAA modes include 8×, 16×, and 16×Q. Each CSAA
mode enhances built-in application antialiasing modes with much higher-quality
antialiasing. The fourth new mode, 8×Q, is standard 8× multisampling.
Each new AA mode can be enabled from the NVIDIA driver control panel and
requires the user to select an option called “Enhance the Application Setting.” Users
must first turn on any antialiasing level within the game’s control panel for the new
AA modes to work, since they need the game to properly allocate and enable
antialiased rendering surfaces. If a game does not natively support antialiasing, a user
can select an NVIDIA driver control panel option called “Override Any
Applications Setting,” which allows any control panel AA settings to be used with
the game. Note that this setting does not always work with all applications. For

TB-02787-001_v1.0 27
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

games with AA capability built in, the “Enhancing the Application Setting” is an
easy way to improve overall image quality.
In many games, the new 16× high quality mode will yield frame-per-second
performance results similar to standard 4× multisampled mode, but with much
improved image quality. In certain cases, such as the edge of stencil shadow
volumes, the new antialiasing modes will not be enabled and those portions of the
scene will fall back to 4× multisampled mode.
Below is a close-up image of how the new CSAA 16× mode compares to standard
4× multisampled AA mode.

Figure 19. Coverage sampling antialiasing (4× MSAA vs. 16× CSAA)

GeForce 8800 GPUs support both the FP16 and the FP32 component for HDR
rendering, which can work simultaneously with multisampled antialiasing delivering
incredibly rich images and scenery.

28 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

Anisotropic Filtering (AF) improves the clarity and sharpness of various scene
objects that are viewed at sharp angles and/or recede into the distance (Figure 20).
One example is a roadway billboard with text that looks skewed and blurred when
viewed at a sharp angle (with respect to the camera) when standard bilinear and
trilinear isotropic texture filtering methods are applied. Anisotropic filtering
(combined with trilinear mipmapping) allows the skewed text to look much sharper.
Similarly, a cobblestone roadway that fades into the distance can be sharpened with
anisotropic filtering.

(Photo courtesy of Wikipedia for public domain use)

Figure 20. Isotropic trilinear mipmapping (left) vs. anisotropic

trilinear mipmapping (right)

Anisotropic filtering is memory bandwidth intensive, particularly at high AF levels.

For example, 16×AF means that up to 16 bilinear reads per each of two adjacent
mipmap levels (128 memory reads total) are performed, and a weighted average is
used to derive final texture color to apply to a pixel.
GeForce 8800 GPU’s new default high-quality anisotropic filtering removes angle
optimizations and yields near perfect anisotropic filtering, as shown in Figure 21.

TB-02787-001_v1.0 29
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Figure 21. Anisotropic filtering comparison

(GeForce 7 Series on left, and GeForce 8 Series or right using
default anisotropic Texture Filtering)

Refer to Lumenex Engine: The New Standard in GPU Image Quality (TB-02824-001) for
more details.
Following is a discussion of three important architectural enhancements that permit
better overall GPU performance.

30 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

Decoupled Shader/Math,
Branching, and Early-Z
Decoupled Shader Math and Texture Operations
Texture addressing, fetching, and filtering can take many GPU core clock cycles. If
an architecture requires a texture to be fetched and filtered before performing the
next math operation in a particular shader, the considerable texture fetch and
filtering (such as 16× anisotropic filtering) latencies can really slow down a GPU.
GeForce 8800 GPUs can do a great job tolerating and essentially “hiding” texture
fetch latency by performing a number of useful independent math operations
concurrently.
For comparison, a GeForce 7 Series GPU texture address calculation was
interleaved with shader floating-point math operations in Shader Unit 1 of a pixel
pipeline. Although this design was chosen to optimize die size, power, and
performance, it could cause some shader math bottlenecks when textures were
fetched, preventing use of a shader processor until the texture was retrieved.
GeForce 8800 GPUs attacks the shader math and texture processing efficiency
problem by decoupling shader and texture operations so that texture operations can
be performed independent of shader math operations.
Figure 22 illustrates math operations (not dependent on specific texture data being
fetched) that can be executed while one or more textures are being fetched from
frame buffer memory, or in the worst case, from system memory. Shader processor
utilization is improved. In effect, while a thread fetching a texture is executing, the
GeForce 8800 GPU’s GigaThread technology can swap in other threads to execute,
ensuring that shader processors are never idle when other work needs to done.

Figure 22. Decoupling texture and math operations

TB-02787-001_v1.0 31
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Branching Efficiency Improvements

An important aspect of overall GPU performance in processing complex DirectX
10 shader workloads is branch efficiency. For background and comparison,
GeForce 7 Series GPUs were designed to be efficient when processing typical
DirectX 9 shaders. When an if-then-else statement was encountered in pixel shader
code, a batch of 880 pixels was processed at once. Some of the pixels in the batch
would generally require pixel shading effects applied based on the “then” code path,
while the other pixels in the batch just went along the same code path for the ride,
but were effectively masked out of any operations. Then the whole batch would take
the “else” code path, where just the opposite would occur, and the other set of
pixels would respond to the “else” code, while the rest went along for the ride.
GeForce 8800 Series GPUs are designed to process complex DX10 shaders.
Programmers will enjoy as fine 16-pixel branching granularity up to 32 pixels in
some cases. Compared to the ATI X1900 series, which uses 48 pixel granularity, the
GeForce 8800 architecture is far more efficient with 32 pixel granularity for pixel
shader programs. The chart in Figure 23 shows perfect branch efficiency for even
numbers of coherent 4×4 pixel tiles.

Figure 23. GeForce 8800 GPU pixel shader branching efficiency

32 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 Architecture in Detail

Early-Z Comparison Checking

Modern GPUs use a Z-buffer (also known as depth buffer) to track which pixels in
a scene are visible to the eye, and which do not need to be displayed because they
are occluded by other pixels. Every pixel has corresponding Z information in the Z-
buffer.
For background, a single 3D frame is processed and converted to a 2D image for
display on a monitor. The frame is constructed from a sequential stream of vertices
sent from the host to the GPU. Polygons are assembled from the vertex stream, and
2D screen-space pixels are generated and rendered.
In the course of constructing a single 2D frame in a given unit of time, such as
1/60th of s second, multiple polygons and their corresponding pixels may overlay
the same 2D screen-based pixel locations. This is often called depth complexity, and
modern games might have depth complexities of three or four, where three or four
pixels rendered in a frame overlay the same 2D screen location.
Imagine polygons (and resulting pixels) for a wall being processed first in the flow
of vertices to build a scene. Next, polygons and pixels for a chair in front of the wall
are processed. For a given 2D pixel location onscreen, only one of the pixels can be
visible to the viewer—a pixel for the chair or a pixel for the wall. The chair is closer
to the viewer, so its pixels are displayed. (Note that some objects may be
transparent, and pixels for transparent objects can be blended with opaque or
transparent pixels already in the background or with pixels already in the frame
buffer from a prior frame).
Figure 24 shows a simple Z-buffering example for a single pixel location. Note that
we did not include actual Z-buffer data in the Z-buffer location.

Figure 24. Example of Z-buffering

TB-02787-001_v1.0 33
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

A few methods use Z-buffer information to help cull or prevent pixels from being
rendered if they are occluded. Z-cull is a method to remove pixels from the pipeline
during the rasterization stage, and can examine and remove groups of occluded
pixels very swiftly.
A GeForce 8800 GTX GPU can cull pixels at four times the speed of GeForce
7900 GTX, but neither GPU catches all occlusion situations at the individual pixel
level.
Z comparisons for individual pixel data have generally occurred late in the graphics
pipeline in the ROP (raster operations) unit. The problem with evaluating individual
pixels in the ROP is that pixels must traverse nearly the entire pipeline to ultimately
discover some are occluded and will be discarded. With complex shader programs
that have hundreds or thousands of processing steps, all the processing is wasted on
pixels that will never be displayed!
What if an Early-Z technique could be employed to test Z values of pixels before
they entered the pixel shading pipeline? Much useless work could be avoided,
improving performance and conserving power.
GeForce 8800 Series GPUs implement an Early-Z technology, as depicted in Figure
25, to increase performance noticeably.

Figure 25. Example of early-Z technology

Next, we’ll look at how the GeForce 8800 GPU architecture redefines the classic
GPU pipeline and implements DirectX 10–compatible features. Later in this
document, we describe key DirectX 10 features in more detail.

34 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 GTX GPU
Design and Performance

W have already covered a lot of the basics, so now we can look at the specifics of
the GeForce 8800 GTX architecture without intimidation. The block diagram
shown in Figure 26 should now look less threatening if you’ve read the prior
sections.

Figure 26. GeForce 8800 GTX block diagram

TB-02787-001_v1.0 35
November 4, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Host Interface and Stream

Processors
Starting from the top of Figure 26 you see the Host interface block, which includes
buffers to receive commands, vertex data, and textures sent to the GPU from the
graphics driver running on the CPU. Next is the input assembler, which gathers
vertex data from buffers and converts to FP 32 format, while also generating
various index IDs that are helpful for performing various repeated operations on
vertices and primitives, and for enabling instancing.
The GeForce 8800 GTX GPU includes 128 efficient stream processors (SPs)
depicted in the diagram, and each SP is able to be assigned to any specific shader
operation. We’ve covered a lot about unified shader architecture and stream
processor characteristics earlier in this paper, so now you can better understand how
the SPs are grouped inside the GeForce 8800 GTX chip. This grouping allows the
most efficient mapping of resources to the processors, such as the L1 caches and
the texture filtering units.
Data can be moved quickly from the output of a stream processor to the input of
another stream processor. For example, vertex data processed and output by a
stream processor can be routed as input to the Geometry Thread issue logic very
rapidly.
It’s no secret that shader complexity and program length continue to grow at a rapid
rate. Many game developers are taking advantage of new DirectX 10 API features
such as stream output, geometry shaders, and improved instancing supported by
GeForce 8800 GPU architecture. These features will add richness to their 3D game
worlds and characters, while shifting more of the graphics and physics processing
burden to the GPU, allowing the CPU to perform more artificial intelligence (AI)
processing.

Raw Processing and Texturing Filtering Power

Each stream processor on a GeForce 8800 GTX operates at 1.35 GHz and supports
the dual issue of a scalar MAD and a scalar MUL operation, for a total of roughly
520 gigaflops of raw shader horsepower. But raw gigaflops do not tell the whole
performance story. Instruction issue is 100 percent efficient with scalar shader units,
and the mixed scalar and vector shader program code will perform much better
compared to vector-based GPU hardware shader units that have instruction issue
limitations (such as 3+1 and 2+2).
Texture filtering units are fully decoupled from the stream processors and deliver 64
pixels per clock worth of raw texture filtering horsepower (versus 24 pixels in the
GeForce 7900 GTX); 32 pixels per clock worth of texture addressing; 32 pixels per
clock of 2× anisotropic filtering; and 32-bilinear-filtered pixels per clock.

36 TB-02787-001_v1.0
November 8, 2006
 GeForce 8800 GTX GPU
Design and Performance

In essence, full-speed bilinear anisotropic filtering is nearly free on GeForce 8800

GPUs. FP16 bilinear texture filtering is also performed at 32 pixels per clock (about
5× faster than GeForce 7x GPUs), and FP16 2:1 anisotropic filtering is done at 16
pixels per clock. Note that the texture units run at the core clock, which is 575 MHz
on the GeForce 8800 GTX.
At the core clock rate of 575 MHz, texture fill rate for both bilinear filtered texels
and 2:1 bilinear anisotropic filtered texels is 575 MHz × 32 = 18.4 billion
texels/second. However, 2:1 bilinear anisotropic filtering uses two bilinear samples
to derive a final filtered texel to apply to a pixel. Therefore, GeForce 8800 GPUs
have an effective 36.8 billion texel/second fill rate when equated to raw bilinear
texture filtering horsepower.
You can see the tremendous improvement of the GeForce 8800 GTX over the
GeForce 7900 GTX in relative filtering speed in Figure 27.

Figure 27. Texture fill performance of GeForce 8800 GTX

ROP and Memory Subsystems

The GeForce 8800 GTX has six Raster Operation (ROP) partitions, and each
partition can process 4 pixels (16 subpixel samples, as shown in the diagram) for a
total of 24 pixel/clock output capability with color and Z processing. For Z-only
processing, an advanced new technique allows up to 192 samples/clock to be
processed when a single sample is used per pixel. If 4× multisampled antialiasing is
enabled, then 48 pixels per clock Z-only processing is possible.
The GeForce 8800 ROP subsystem supports multisampled, supersampled, and
transparency adaptive antialiasing. Most important is the addition of four new
single-GPU antialiasing modes (8×, 8×Q, 16×, and 16×Q), which provide the
absolute best antialiasing quality for a single GPU in the market today.

TB-02787-001_v1.0 37
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

The ROPs also support frame buffer blending of FP16 AND FP32 render targets,
and either type of FP surface can be used in conjunction with multisampled
antialiasing for outstanding HDR rendering quality. Eight MRTs (Multiple Render
Targets) can be utilized, which is also supported by DX10. Each of the MRTs can
define different color formats. New high-performance, more efficient compression
technology is implemented in the ROP subsystem to accelerate color and Z
processing.
As shown in Figure 26 six memory partitions exist on a GeForce 8800 GTX GPU,
and each partition provides a 64-bit interface to memory, yielding a 384-bit
combined interface width. The 768 MB memory subsystem implements a high-
speed crossbar design, similar to GeForce 7x GPUs, and supports DDR1, DDR2,
DDR3, GDDR3, and GDDR4 memory. The GeForce 8800 GTX uses GDDR3
memory default clocked at 900 MHz. With a 384-bit (48 byte wide) memory
interface running at 900 MHz (1800 MHz DDR data rate), frame buffer memory
bandwidth is very high at 86.4 GBps. With 768 MB of frame buffer memory, far
more complex models and textures can be supported at high resolutions and image
quality settings.

Balanced Architecture
NVIDIA engineers spent a great deal of time ensuring the GeForce 8800 GPU
Series was a balanced architecture. It wouldn’t make sense to have 128 streaming
processors or 64 pixels worth of texture filtering power if the memory subsystem
weren’t able to deliver enough data, or if the ROPs were a bottleneck processing
pixels, or if the clocking of different subsystems was mismatched. Also, the GPUs
must be built in a manner that makes them power efficient and die-size efficient
with optimal performance. The graphics board must be able to be integrated into
mainstream computing systems without extravagant power and cooling.
Each of the unified streaming processors can handle different types of shader
programs to allow instantaneous balancing of processor resources based on
demand. Internal caches are designed for extremely high performance and hit rates,
and combined with the high-speed and large frame buffer memory subsystem, the
streaming processors are not starved for data.
During periods of texture fetch and filtering latency, GigaThread technology can
immediately dispatch useful work to a processor that, in past architectures, may have
needed to wait for the texture operation to complete. With more vertex and pixel
shader program complexity, many more cycles will be spent processing in the shader
complex, and the ROP subsystem capacity was built to be balanced with shader
processor output. And the 900 MHz memory subsystem ensures even the highest-
end resolutions with high-quality filtering can be processed effectively.
We have talked a lot about hardware and cannot forget that drivers play a large part
in balancing overall performance. NVIDIA ForceWare® drivers work hand-in-hand
with the GPU to ensure superior GPU utilization with minimal CPU impact.
Now that you have a good understanding of GeForce 8800 GPU architecture, let’s
look at DirectX 10 features in more detail. You will then be able to relate the
DirectX 10 pipeline improvements to the GeForce 8800 GPU architecture.

38 TB-02787-001_v1.0
November 8, 2006
 DirectX 10 Pipeline

The DirectX 10 specification, combined with DX10-capable hardware, has relieved

many of the constraints and problems of pre-DirectX 10 classic graphics pipelines.
In addition to a new unified instruction set and increases in resources, two of the
more visible additions are an entirely new programmable pipeline stage called the
geometry shader, and the stream output feature.
DirectX 10 and prior versions (with programmable pipeline capabilities) were
designed to operate like a virtual machine, where the GPU is virtualized, and device-
independent shader code is compiled to specific GPU machine code at runtime by
the GPU driver’s built-in Just-In-Time (JIT) compiler.
DirectX 9 Shader Model 3 used different virtual machine models with different
instructions and different resources for each of the programmable pipeline stages.
DirectX 10’s Shader Model 4 virtual machine model provides a “common core” of
resources for each programmable shader stage (vertex, pixel, geometry) with many
more hardware resources available to shader programs. Let’s look at the new
virtualization model and Shader Model 4 a bit more closely.

Virtualization and Shader

Model 4
You are likely familiar with the concept of virtualization of computing resources,
such as virtual memory, virtual machines (Java VMs, for example), virtual I/O
resources, operating system virtualization, and so forth.
Direct X shader assembly language is similar to Java VM language because both are
a machine-independent intermediate language (IL) compiled to a specific machine
language by a Just-In-Time (JIT) compiler. As mentioned above, shader assembly
code gets converted at runtime by the GPU driver into GPU-specific machine
instructions using a JIT complier built into the driver. (Microsoft high-level shader
language (HLSL) and NVIDIA Cg high-level shader programming languages both
get compiled down to the shader assembly IL format.)

TB-02787-001_v1.0 39
November 4, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

While similar in many respects to Shader Model 3, new features added with Shader
Model 4 include a new unified instruction set; many more registers and constants;
integer computation; unlimited program length; fewer state changes (less CPU
intervention); 8 multiple render target regions instead of 4; more flexible vertex
input via the input assembler; the ability of all pipeline stages to access buffers,
textures, and render targets with few restrictions; and the capability of data to be
recirculated through pipeline stages (stream out).
Shader Model 4 also includes a very different render state model, where application
state is batched more efficiently, and more work can be pushed to the GPU with
less CPU involvement. Table 1 shows DirectX 10 Shader Model 4 versus prior
shader models.

Table 1. Shader Model Progression

DX8 SM1.X DX9 SM2 DX9 SM3 DX10
Vertex instructions 128 256 512 64 K
Pixel instructions 4+8 32+64 512
Vertex constants 96 256 256 16 × 4096
Pixel constants 8 32 224
Vertex temps 16 16 16 4096
Pixel temps 2 12 32
Vertex inputs 16 16 16 16
Pixel inputs 4+2 8+2 10 32
Render targets 1 4 4 8
Vertex textures N/A N/A 4 128
Pixel textures 8 16 16
2D tex size – – 2K×2K 8K×8K
Int ops – – – Yes
Load ops – – – Yes
Derivatives – – Yes Yes
Vertex flow control N/A Static Static/Dyn Dynamic
Pixel flow control N/A N/A Static/Dyn

40 TB-02787-001_v1.0
November 8, 2006
 DirectX 10 Pipeline

Stream Output
Stream output is a very important and useful new DirectX 10 feature supported in
GeForce 8800 GPUs. Stream output enables data generated from geometry shaders
(or vertex shaders if geometry shaders are not used) to be sent to memory buffers
and subsequently forwarded back into the top of the GPU pipeline to be processed
again (Figure 28). Such dataflow permits more complex geometry processing,
advanced lighting calculations, and GPU-based physical simulations with little CPU
involvement.

Figure 28. Direct3D 10 pipeline

Stream output is a more generalized version of the older “render to vertex buffer”
feature that permits data generated from geometry shaders (or from vertex shaders
if geometry shaders are not used) to be sent to “stream buffers” and subsequently
forwarded back to the top of the pipeline to be processed again. (See “The Hair
Challenge” for an example of usage.)

TB-02787-001_v1.0 41
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Geometry Shaders
High polygon–count characters with realistic animation and facial expressions are
now possible with DirectX 10 geometry shading, as are natural shadow volumes,
physical simulations, faster character skinning, and a variety of other geometry
operations.
Geometry shaders can process entire primitives as inputs and generate entire
primitives as output, rather than processing just one vertex at a time, as with a
vertex shader. Input primitives can be comprised of multiple vertices, such as point
lists, line lists or strips, triangle lists or strips, a line list or strip with adjacency info,
or a triangle list or strip with adjacency info. Output primitives can be point lists,
line strips, or triangle strips.
Limited forms of tessellation—breaking down primitives such as triangles into a
number of smaller triangles to permit smoother edges and more detailed objects—
are possible with geometry shaders. Examples could include tessellation of water
surfaces, point sprites, fins, and shells. Geometry shaders can also control objects
and create and destroy geometry (they can read a primitive in and generate more
primitives, or not emit any primitives as output).
Geometry shaders also can extrude silhouette edges, expand points, assist with
render to cube maps, render multiple shadow maps, perform character skinning
operations, and enable complex physics and hair simulations. And, among other
things, they can generate single-pass environment maps, motion blur, and stencil
shadow polygons, plus enable fully GPU-based particle systems with random
variations in position, velocity, and particle lifespan.
Note that software-based rendering techniques in existence for years can provide
many of these capabilities, but they are much slower, and this is the first time such
geometry processing features are implemented in the hardware 3D pipeline.
A key advantage of hardware-based geometry shading is the ability to move certain
geometry processing functions from the CPU to the GPU for much better
performance. Characters can be animated without having the CPU intervene, and
true displacement mapping is possible, permitting vertices to be moved around to
create undulating surfacing and other cool effects.

42 TB-02787-001_v1.0
November 8, 2006
 DirectX 10 Pipeline

Improved Instancing
DirectX 9 introduced the concept of object instancing, where a single API draw call
would send a single object to the GPU, followed by a small amount of “instance
data” that can vary object attributes such as position and color. By applying the
varying attributes to the original object, tens or hundreds of variations of an object
could be created without CPU involvement (such as leaves on tree or an army of
soldiers).
DirectX 10 adds much more powerful instancing by permitting index values of
texture arrays, render targets, and even indices for different shader programs to be
used as the instance data that can vary attributes of the original object to create
different-looking versions of the object. And it does all this with fewer state changes
and less CPU intervention.

Figure 29. Instancing at work—numerous characters rendered

In general, GeForce 8800 Series GPUs work with the DX10 API to provide
extremely efficient instancing and batch processing of game objects and data to
allow for richer and more immersive game environments.

TB-02787-001_v1.0 43
November 8, 2006
NVIDIA GeForce 8800 Architecture Technical Brief

Vertex Texturing
Vertex texturing was possible in DirectX 9 and is now a major feature of the
DirectX 10 API and able to be used with both vertex shaders and geometry shaders.
With vertex texturing, displacement maps or height fields are read from memory
and their “texels” are actually displacement (or height) values, rather than color
values. The displacements are used to modify vertex positions of objects, creating
new shapes, forms, and geometry-based animations.

The Hair Challenge

The Broadway musical Hair said it best: “Long, straight, curly, fuzzy, snaggy, shaggy,
ratty, matty, oily, greasy, fleecy, shining, streaming, flaxen, waxen, knotted, polka-
dotted, twisted, beaded, braided, powdered, flowered, confettied, bangled, tangled,
spangled, spaghettied, and a real pain to render realistically in a 3D game.” OK,
maybe not the last clause, but it’s true!
A good example of the benefit of DirectX 10 and GeForce 8800 GPUs is in
creating and animating complex and realistic-looking hair. Rendering natural-looking
hair is both a challenging rendering problem and a difficult physics simulation
problem.

Table 2. Hair Algorithm Comparison of DirectX 9 and DirectX 10

Algorithm GeForce 7 Series GeForce 8 Series
Physical simulation on control points CPU GPU
Interpolate and tessellate control points CPU GPU – GS
Save tessellated hairs to memory CPU GPU – SO
Render hair to deep shadow map GPU GPU
Render hair to back buffer GPU GPU

With DirectX 9, the physics simulation of the hair is performed on the CPU.
Interpolation and tessellation of the control points of the individual hairs in the
physics simulation is also performed by the CPU. Next, the hairs must be written to
memory and copied to the GPU, where they can finally be rendered. The reason we
don’t see very realistic hair with DirectX 9 games is that it’s simply too CPU-
intensive to create, and developers can’t afford to spend huge amounts of CPU
cycles just creating and animating hair at the expense of other more important game
play objects and functions.
With DirectX 10, the physics simulation of the hair is performed on the GPU, and
interpolation and tessellation of control points is performed by the geometry shader.
The output from the geometry shader is transferred to memory using stream output,
and read back into the pipeline to actually render the hair.
Expect to see far more realistic hair in DX10 games that take advantage of the
power of GeForce 8800 Series GPUs.

44 TB-02787-001_v1.0
November 8, 2006
 Conclusion

As you are now aware, the GeForce 8800 GPU architecture is a radical departure
from prior GPU designs. Its massively parallel unified shader design delivers
tremendous processing horsepower for high-end 3D gaming at extreme resolutions,
with all quality knobs set to the max. New antialiasing technology permits 16× AA
quality at the performance of 4× multisampling, and 128-bit HDR rendering is now
available and can be used in conjunction with antialiasing.
Full DirectX10 compatibility with hardware implementations of geometry shaders,
stream out, improved instancing, and Shader Model 4 assure users they can run their
DirectX 10 titles with high performance and image quality. All DirectX 9, OpenGL,
and prior DirectX titles are fully compatible with the GeForce 8800 GPU unified
design and will attain the best performance possible.
PureVideo functionality built in to all GeForce 8800–class GPUs ensures flawless
SD and HD video playback with minimal CPU utilization. Efficient power
utilization and management delivers outstanding performance per watt and
performance per square millimeter.
Teraflops of floating-point processing power, SLI capability, support for thousands
of threads in flight, Early-Z, decoupled shader and math processing, high-quality
anisotropic filtering, significantly increased texture filtering horsepower and memory
bandwidth, fine levels of branching granularity, plus the 10-bit display pipeline and
PureVideo feature set—all these features contribute to making the GeForce 8800
GPU Series the best GPU architecture for 3D gaming and video playback
developed to date.

TB-02787-001_v1.0 45
November 4, 2006
Notice
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND
OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA
MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE
MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT,
MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no
responsibility for the consequences of use of such information or for any infringement of patents or other
rights of third parties that may result from its use. No license is granted by implication or otherwise under any
patent or patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to
change without notice. This publication supersedes and replaces all information previously supplied. NVIDIA
Corporation products are not authorized for use as critical components in life support devices or systems
without express written approval of NVIDIA Corporation.
Trademarks
NVIDIA, the NVIDIA logo, CUDA, ForceWare, GeForce, GigaThread, Lumenex, NVIDIA nForce, PureVideo, SLI,
and Quantum Effects are trademarks or registered trademarks of NVIDIA Corporation in the United States
and other countries. Other company and product names may be trademarks of the respective companies
with which they are associated
Copyright
© 2006 NVIDIA Corporation. All rights reserved.

NVIDIA Corporation
2701 San Tomas Expressway
Santa Clara, CA 95050
www.nvidia.com