Tim Sweeney

CEO, Founder
Epic Games
[email protected]

THE END OF THE GPU

ROADMAP
Background:
Epic Games
Background: Epic Games
 Independent game developer
 Located in Raleigh, North Carolina, USA
 Founded in 1991
 Over 30 games released
 Gears of War
 Unreal series
 Unreal Engine 3 is used by 100’s of games
 History:
Unreal Engine
 Unreal Engine 1
1996-1999
 First modern game engine
 Object-oriented
 Real-time, visual toolset
 Scripting language
 Last major software renderer
 Software texture mapping
 Colored lighting, shadowing
 Volumetric lighting & fog
 Pixel-accurate culling
 25 games shipped
 Unreal Engine 2
2000-2005

 PlayStation 2, Xbox, PC
 DirectX 7 graphics
 Single-threaded
 40 games shipped
 Unreal Engine 3
2006-2012
 PlayStation 3, Xbox 360, PC
 DirectX 9 graphics
 Pixel shaders
 Advanced lighting & shadowing
 Multithreading (6 threads)
 Advanced physics
 More visual tools
 Game Scripting
 Materials
 Animation
 Cinematics…
 150 games in
development
 Unreal Engine 3 Games

Army of Two (Electronic Arts)

Mass Effect (BioWare)

Undertow (Chair Entertainment)

BioShock (2K Games)
Game Development: 2009
 Gears of War 2: Project Overview
 Project Resources
 15 programmers
 45 artists
 2-year schedule
 $12M development budget
 Software Dependencies
 1 middleware game engine
 ~20 middleware libraries
 Platform libraries
 Gears of War 2: Software Dependencies

Gears of War 2
Gameplay Code
~250,000 lines C++, script code

Unreal Engine 3
Middleware Game Engine
~2,000,000 lines C++ code

ZLib
DirectX OpenAL Speed
Tree
FaceFX
Face
Bink
Movie
Data …
Graphics Audio Rendering Animation Codec
Compr-
ession
Hardware:
History
Computing History

1985 Intel 80386: Scalar, in-order CPU

1989 Intel 80486: Caches!
1993 Pentium: Superscalar execution
1995 Pentium Pro: Out-of-order execution
1999 Pentium 3: Vector floating-point
2003 AMD Opteron: Multi-core
2006 PlayStation 3, Xbox 360: “Many-core”
…and we’re back to in-order execution
Graphics History

1984 3D workstation (SGI)

1997 GPU (3dfx)
2002 DirectX9, Pixel shaders (ATI)
2006 GPU with full programming language
(NVIDIA GeForce 8)
2009? x86 CPU/GPU Hybrid
(Intel Larrabee)
Hardware:
2012-2020
 Hardware: 2012-2020
Processor Processor Processor Processor Processor
In Order In Order In Order In Order In Order
4 Threads 4 Threads 4 Threads 4 Threads 4 Threads
I$ D$ I$ D$ I$ D$ I$ D$ I$ D$

L2 Cache

Processor Processor Processor Processor Processor

In Order In Order In Order In Order In Order
4 Threads 4 Threads 4 Threads 4 Threads 4 Threads
I$ D$ I$ D$ I$ D$ I$ D$ I$ D$

Intel Larrabee NVIDIA GeForce 8

 x86 CPU-GPU Hybrid  General Purpose GPU
 C/C++ Compiler  CUDA “C” Compiler
 DirectX/OpenGL  DirectX/OpenGL
 Many-core, vector architecture  Many-core, vector architecture
 Teraflop-class performance  Teraflop-class performance
Hardware: 2012-2020

CONCLUSION
CPU, GPU architectures are getting closer
THE GPU TODAY
 The GPU Today

 Large frame buffer

 Complicated pipeline
 It’s fixed-function
 But we can specify
shader programs
that execute in certain pipeline stages
Shader Program Limitations

 No random-access memory writes

 Can write to current pixel in frame buffer
 Can’t create data structures
 Can’t traverse data structures
 Can hack it using texture accesses
 Hard to share data between main program
and shaders programs
 Weird programming language
 HLSL rather than C/C++

Result: “The Shader ALU Plateau”

Antialiasing Limitations

 MSAA & Oversampling

 Every 1 bit of output precision costs up to
2X memory & performance!
 Ideally want 10-20 bits
 Discrete sampling (in general)
 Texture filtering only implies antialiasing when
shader equation is linear
 Most shader equations are nonlinear

Aliasing is the #1 visual artifact in Gears of War

Texture Sampling Limitations

 Inherent artifacts of bilinear/trilinear

 Poor approximation of Integrate(color,area)
in the presence of:
 Small triangles
 Texture seams
 Alpha translucency
 Masking
 Fixed-function = poor scalability
 Megatexture, etc
Frame Buffer Model Limitation

 Frame buffer: 1 (or n) layers of 4-vectors,

where n = small constant
 Ineffective for
 General translucency
 Complex shadowing models
 Memory bandwidth requirement =
FPS * Pixel Count * Layers Depth * pow(2,n)
where n = quality of MSAA
Summary of Limitations

 “The Shader ALU Plateau”

 Antialiasing limitations
 Texture Sampling limitations
 Frame Buffer limitations
The Meta-Problem:

 The fixed-function pipeline is

too fixed to solve its problems
 Result:
 All games look similar
 Derive little benefit from Moore’s Law
 Crysis on high-end NVIDIA SLI solution only looks at
most marginally better than top Xbox 360 games

This is a market BEGGING

to be disrupted :-)
SO...
Return to 100% “Software” Rendering

 Bypass the OpenGL/DirectX API

 Implement a 100% software renderer
 Bypass all fixed-function pipeline hardware
 Generate image directly
 Build & traverse complex data structures
 Unlimited possibilities

Could implement this…

 On Intel CPU using C/C++
 On NVIDIA GPU using CUDA (no DirectX)
Software Rendering in Unreal 1 (1998)

Ran 100% on CPU

No GPU required!

Features
 Real-time colored lighting
 Volumetric Fog
 Tiled Rendering
 Occlusion Detection
 Software Rendering in 1998 vs 2012

60 MHz Pentium could execute:

16 operations per pixel
at 320x200, 30 Hz

In 2012, a 4 Teraflop processor

would execute:
16000 operations per pixel
at 1920x1080, 60 Hz

Assumption: Using 50% of computing power for graphics, 50% for gameplay
Future Graphics:
Raytracing

 For each pixel

 Cast a ray off into scene
 Determine which objects were hit
 Continue for reflections, refraction, etc

 Consider
 Less efficient than pure rendering
 Can use for reflections in traditional render
Future Graphics:
The REYES Rendering Model
 “Dice” all objects in scene down into sub-pixel-
sized triangles
 Rendering with
 Flat Shading (!)
 Analytic antialiasing
 Per-pixel occlusion
(A-Buffer/BSP)
 Benefits
 Displacement maps for free
 Analytic Antialiasing
 Advanced filtering (Gaussian)
 Eliminates texture sampling
 Future Graphics:
The REYES Rendering Model

Today’s Pipeline Potential 2012 Pipeline

 Build 4M poly “high-res” character  Build 4M poly “high-res” character
 Generate normal maps from  Render it in-game!
geometry in high-res  Advanced LOD scheme assures
 Rendering 20K poly “low-res” proper sub-pixel sized triangles
character in-game
Future Graphics:
Volumetric Rendering
 Direct Voxel Rendering
 Raycasting
 Efficient for trees, foliage
 Tesselated Volume Rendering
 Marching Cubes
 Marching Tetrahedrons
 Point Clouds
 Signal-Space Volume Rendering
 Fourier Projection Slice Theorem
 Great for clouds, translucent volumetric data
Future Graphics:
Software Tiled Rendering

 Split the frame buffer up into bins

 Example: 1 bin = 8x8 pixels
 Process one bin at a time
 Transform, rasterize all objects in the bin

 Consider
 Cache efficiency
 Deep frame buffers, antialiasing
Hybrid Graphics Algorithms

 Analytic Antialiasing
– Analytic solution, better than 1024x MSAA

 Sort-independent translucency
– Sorted linked-list per pixel of fragments requiring per-pixel memory
allocation, pointer-following, conditional branching (A-Buffer).

 Advanced shadowing techniques

– Physically accurate per-pixel penumbra volumes
– Extension of well-known stencil buffering algorithm
– Requires storing, traversing, and updating a very simple BSP tree per-
pixel with memory allocation and pointed following.

 Scenes with very large numbers of objects

– Fixed-function GPU + API has 10X-100X state change disadvantage
 Graphics: 2012-2020
Potential Industry Goals

Achieve movie-quality:
 Antialiasing
 Direct Lighting
 Shadowing
 Particle Effects
 Reflections

Significantly improve:
 Character animation
 Object counts
 Indirect lighting
SOFTWARE IMPLICATIONS
 Software Implications

Software must scale to…

• 10’s – 100’s of threads
• Vector instruction sets
 Software Implications

Programming Models
• Shared State Concurrency
• Message Passing
• Pure Functional Programming
• Software Transactional Memory
Multithreading in Unreal Engine 3:
“Task Parallelism”
 Gameplay thread
 AI, scripting
 Thousands of interacting objects

 Rendering thread
 Scene traversal, occlusion
 Direct3D command submission

 Pool of helper threads for other work

 Physics Solver
 Animation Updates

Good for 4 threads.

No good for 100 threads!
“Shared State Concurrency”
The standard C++/Java threading model

 Many threads are running

 There is 512MB of data
 Any thread can modify any data at any time
 All synchronization is explicit, manual
 See: LOCK, MUTEX, SEMAPHORE
 No compile-time verification of correctness properties:
 Deadlock-free
 Race-free
 Invariants
 Multithreaded Gameplay Simulation:
Manual Synchronization

Idea:
 Update objects in multiple threads
 Each object contains a lock
 “Just lock an object before using it”

Problems:
 “Deadlocks”
 “Data Races”
 Debugging is difficult/expensive
Multithreaded Gameplay Simulation:
“Message Passing”

Idea:
 Update objects in multiple threads
 Each object can only modify itself
 Communicate with other objects by sending
messages

Problems:
 Requires writing 1000’s of message protocols
 Still need synchronization
 Pure Functional Programming

“Pure Functional” programming style:

• Define algorithms that don’t write to shared
memory or perform I/O operations

(their only effect is to return a result)

Examples:
• Collision Detection
• Physics Solver
• Pixel Shading
 Pure Functional Programming

“Inside a function with no side effects,

sub-computations can be run in any order,
or concurrently,
without affecting the function’s result”

With this property:

• A programmer can explicitly multithread the
code, safely.
• Future compilers will be able to automatically
multithread the code, safely.

See: “Implementing Lazy Functional Languages on Stock Hardware”;

Simon Peyton Jones; Journal of Functional Programming 2005
 Multithreaded Gameplay Simulation:
Software Transactional Memory
Idea:
 Update objects in multiple threads
 Each thread runs inside a transaction block
and has an atomic view of its “local” changes to memory
 C++ runtime detects conflicts between transactions
 Non-conflicting transactions are applied to “global” memory
 Conflicting transactions are “rolled back” and re-run
Implemented 100% in software; no custom hardware required.

Problems:
 “Object update” code must be free of side-effects
 Requires C++ runtime support
 Cost around 30% performance

See: “Composable Memory Transactions”; Tim Harris, Simon Marlow, Simon Peyton Jones,
and Maurice Herlihy. ACM Conference on Principles and Practice of Parallel Programming 2005
Vectorization
Supporting “Vector Instruction Sets” efficiently

NVIDIA GeForce 8:
• 8 to 15 cores
• 16-wide vectors
Vectorization

C++, Java compilers generate “scalar” code

GPU Shader compilers generate “vector” code

 Arbitrary vector size (4, 16, 64, …)
 N-wide vectors yield N-wide speedup
Vectorization: “The Old Way”

 “Old Vectors” (SIMD):

Intel SSE, Motorola Altivec
 4-wide vectors
 4-wide arithmetic operations
 Vector loads
Load vector register from vector stored in memory

 Vector swizzle & mask

Future Programming Models:
Vectorization

 “Old Vectors”
Intel SSE, Motorola Altivec
x0 x1 x2 x3
vec4 x,y,z;
... + + + +
z = x+y;
y0 y1 y2 y3

= = = =

z0 z1 z2 z3
 Vectorization: “New Vectors”

(ATI, NVIDIA GeForce 8, Intel Larrabee)

 16-wide vectors
 16-wide arithmetic
 Vector loads/stores
 Load 16-wide vector register from scalars
from 16 independent memory addresses,
where the addresses are stored in a vector!
 Analogy: Register-indexed constant access in DirectX

 Conditional vector masks

“New SIMD” is better than “Old SIMD”

 “Old Vectors” were only useful when dealing

with vector-like data types:
 “XYZW” vectors from graphics
 4x4 matrices

 “New Vectors” are far more powerful:

Any loop whose body has a statically-known call graph
free of sequential dependencies can be “vectorized”,
or compiled into an equivalent 16-wide vector
program. And it runs up to 16X faster!
 “New Vectors” are universal
int n;
cmplx coords[];

(Mandelbrot set generator)

int color[] = new int[n]

for(int i=0; i<n; i++) {

int j=0;
cmplx c=cmplx(0,0)
while(mag(c) < 2) {
c=c*c + coords[i];
j++;
}
color[i] = j;
}

This code…
 is free of sequential dependencies
 has a statically known call graph
Therefore, we can mechanically transform it into an equivalent
data parallel code fragment.
 “New Vectors” Translation
for(int i=0; i<n; i++) { for(int i=0; i<n; i+=N) {
… i_vector={i,i+1,..i+N-1}
} i_mask={i<n,i+1<N,i+2<N,..i+N-1<N}
…
}

Standard data-parallel loop setup

Note: Any code outside this loop

(which invokes the loop)
is necessarily scalar!
 “New Vectors” Translation
int n;
cmplx coords[];
int color[] = new int[n]
Note: Any code outside this loop
for(int i=0; i<n; i++) {
(which invokes the loop)
int j=0;
cmplx c=cmplx(0,0)
is necessarily scalar!
while(mag(c) < 2) {
c=c*c +
coords[i];
j++; int n;
} cmplx coords[];
color[i] = j; int color[] = new int[n]
}
for(int i=0; i<n; i+=N) {
int[N] i_vector={i,i+1,..i+N-1}
Loop Index Vector bool[N] i_mask={i<n,i+1<N,i+2<N,..i+N-1<N}

complx[N] c_vector={cmplx(0,0),..}
Loop Mask Vector
while(1) {
bool[N] while_vector={
Vectorized Loop Variable i_mask[0] && mag(c_vector[0])<2,
..
Vectorized Conditional: }
if(all_false(while_vector))
Propagates loop mask break;
to local condition c_vector=c_vector*c_vector + coords[i..i+N-1 : i_mask]
}
colors[i..i+N-1 : i_mask] = c_vector;
}

Mask-predicated Mask-predicated
vector write vector read
 Vectorization Tricks
 Vectorization of loops
 Subexpressions independent of loop variable are scalar and can be
lifted out of loop
 Subexpressions dependent on loop variable are vectorized
 Each loop iteraction computes an “active mask” enabling operation
on some subset of the N components
 Vectorization of function calls
 For every scalar function, generate an N-wide vector version of the
function taking an N-wide “active mask”
 Vectorization of conditionals
 Evaluate N-wide conditional and combine it with the current active
mask
 Execute “true” branch if any masked conditions true
 Execute “false” branch if any masked conditions false
 Will often execute both branches
Vectorization Paradigms
 Hand-coded vector operations
 Current approach to SSE/Altivec

 Loop vectorization
 See: Vectorizing compilers

 Run a big function with a big bundle of data

 CUDA/OpenCL

 Nested Data Parallelism

 See NESTL
 Very general set of “vectorization” transforms
for many categories of nested computations
 Layers: Multithreading & Vectors

Physics, collision detection, scene

traversal, path finding ..

Game World State

Graphics shader programs

Vector (Data Parallel) Subset

Purely functional core

Software Transactional Memory

Sequential Execution

Hardware I/O
Potential Performance Gains*: 2012-2020
Up to...
 64X for multithreading
 1024X for multithreading + vectors!

1024X
64X

64X

1X

* My estimate of feasibility based on Moore’s Law

Multithreading & Vectorization:
Who Choses?

 Hardware companies impose a limited

model on developers
 Sony Cell, NVIDIA CUDA, Apple OpenCL

 Hardware provides general feature;

languages & runtimes make it nice;
users choose!
 Tradeoffs
 Performance
 Productivity
 Familiarity
HARDWARE IMPLICATIONS
The Graphics Hardware of the Future

All else is just computing!

Future Hardware:
A unified architecture for computing and graphics

Hardware Model
 Three performance dimensions
 Clock rate
 Cores
 Vector width
 Executes two kinds of code:
 Scalar code (like x86, PowerPC)
 Vector code (like GPU shaders or SSE/Altivec)
 Some fixed-function hardware
 Texture sampling
 Rasterization?
Vector Instruction Issues

 A future computing device needs…

 Full vector ISA
 Masking & scatter/gather memory access
 64-bit integer ops & memory addressing
 Full scalar ISA
 Dynamic control-flow is essential
 Efficient support for scalar<->vector transitions
 Initiating a vector computation
 Reducing the results
 Repacking vectors
 Must support billions of transitions per second
Memory System Issues

Effective bandwidth demands will be huge

Typically read 1 byte of memory per FLOP

4 TFLOP of computing power

demands
4 TBPS of effective memory bandwidth!

Yes, really!
 Memory System Issues

Threads (GPU) Caches (CPU)

 Hide memory latency  Expose memory latency
 Lose data locality  Exploit data locality
to minimize main
memory bandwidth
Memory System Issues

 Cache coherency is vital

 It should be the default
Revisiting REYES

 “Dice” all objects in scene down into sub-

pixel-sized triangles
 Tile-based setup

 Rendering with
 Flat Shading
 No texture sampling
 Analytic antialiasing
 Per-pixel occlusion
(A-Buffer/BSP) Requires no artificial
software threading
or pipelining.
LESSONS LEARNED
 Lessons learned:
Productivity is vital!
Hardware will become 20X faster, but:
 Game budgets will increase less than 2X.

Therefore...
 Developers must be willing to sacrifice performance
in order to gain productivity.
 High-level programming beats
low-level programming.
 Easier hardware beats faster hardware!
 We need great tools: compilers, engines, middleware
libraries...
 Lessons learned:
Today’s hardware is too hard!
 If it costs X (time, money, pain) to develop an efficient
single-threaded algorithm, then…
 Multithreaded version costs 2X
 PlayStation 3 Cell version costs 5X
 Current “GPGPU” version is costs: 10X or more

 Over 2X is uneconomical for most software companies!

 This is an argument against:

 Hardware that requires difficult programming techniques
 Non-unified memory architectures
 Limited “GPGPU” programming models
 Lessons learned:
Plan Ahead
Previous Generation:
 Lead-time for engine development was 3 years
 Unreal Engine 3:
 2003: development started
 2006: first game shipped

Next Generation:
 Lead-time for engine development is 5 years
 Start in 2009, ship in 2014!

So, let’s get started!

CONCLUSION
END