Teknik Antarmuka dan Periferal

Sesi 7
TEI302
 What is a GPU?
• It is a GPU (Graphical Processing Unit)
• First introduced in 1999
• First non-graphical applications started 2003
• Had many problems for non-graphical applications
– ie. Programmer needed knowledge of API and
architecture
• Rendering is image generating from a model using computer
program
• Image model is a 3D object in a structured data and language
which contains geometric, viewpoint, texture, lighting and
shading information
 Texture Mapping
• A layering process for a 3D image using specific surface
• 3D image result will look real and lively
• Fundamentally, texture consists of 2D images at specific
pattern
• This texture encloses the 3D object to form a new object
Texture Mapping Example
 Shader
• A program to define the property of a 3D surface
– Light absorption
– Diffusion
– Reflection
– Shading effect
– Uses ‘shading language’
 Rendering
• Final process to display a completed 3D object including all the properties,
i.e. lighting effect, to produce realistic object
• Common use to perform rendering in video gaming, simulator, movie,
special effect, architecture design visualization
 Graphic Card

• An add-on card in a computer to produce pixels for displaying

the images as the product of CPU computation

GPU

Video memory

Cooler (Heatsink+Fan)

Interface : ISA/PCI/PCIX/AGP/PCIe
 GPU : Graphics Pipeline

Graphics State

Screenspace triangles (2D)

Xformed, Lit Vertices (2D)

Final Pixels (Color, Depth)

Fragments (pre-pixels)
Vertices (3D)

Transform Assemble Video

Application Rasterize Shade
& Light Primitives Memory
(Textures)

CPU GPU Render-to-texture

 The Development of GPU: Modern
Graphics Pipeline
Graphics State

Screenspace triangles (2D)

Xformed, Lit Vertices (2D)

Final Pixels (Color, Depth)

Fragments (pre-pixels)
Vertices (3D)

Vertex
Transform Assemble Fragment
Application Rasterize Shade Video
Processor
& Light Primitives Processor Memory
(Textures)

CPU GPU Render-to-texture

• Programmable vertex • Programmable pixel

processor! processor!
 Advancement of GPU : Modern
Graphics Pipeline
Graphics State

Screenspace triangles (2D)

Xformed, Lit Vertices (2D)

Final Pixels (Color, Depth)

Fragments (pre-pixels)
Vertices (3D)

Vertex Geometry
Assemble Fragment
Application Rasterize Video
Processor Processor
Primitives Processor Memory
(Textures)

CPU GPU Render-to-texture

 Programmable primitive  More flexible memory

assembly! access!
GPU : Computation Power
 GPU: Computation Capacity
Why keeps getting faster?
• Naturally the GPU is easier to add transistor for computation
purposes
• The economics: graphic business market, especially in video gaming
is huge

Graphics Characteristics
• Requires massive computation
• Highly parallelism
• Graphics pipeline is designed to reduce interdependent
operations
• GPUs accommodate massive parallelism by having ALUs to process
arithmetic and overcome the issue of data streaming due to huge
quantity of data
General Purpose CPU
GPU
GPU Structure
GPU Structure
GPU vs CPU
 GPGPU (GPU Computing)
• Applications
• Molecular dynamics
• Electromagnetic and acoustic waves
• Computer vision
• Computational statistics
• Computational finance
• Bioinformatics
 CUDA
• CUDA - Compute Unified Device Architecture
• Changed the architecture to better suit general programming
• CUDA is a software and hardware architecture
• Supports C Programming
• Replaced pixel and vertex pipelines with a single pipeline
• Added SIMT (single instruction multiple thread)
• Still programmers asked for more
 Tesselation GPU
Datasets of polygons (called vertex sets) management to
present objects in a scene and divide them into suitable
structures for rendering. Usually in the form of triangles.
 Next Gen Code: “Fermi”
• Improved double precision performance
• ECC Support (error correction code)
• True Cache Hierarchy
• More Shared Memory
• Faster Context Switching
• Faster Atomic Operations
 “Fermi” Architecture
• 16 SMs (streaming multiprocessors)
• Each SM contains 32 CUDA cores
• Totaling 512 CUDA cores
• Has 64-bit memory partitions
• Supports up to 6 GB GDDR5 DRAM
• Connects to a host CPU via PCI Express
 Programming Model
• Parallel functions are done on the GPU
• Non parallel code is done on the host CPU
• Parallel functions are organized in threads/thread
blocks/arrays of thread blocks
• Threads each have it’s own private memory
• Thread blocks share memory with each thread in it’s block
• Arrays of blocks share memory for application
Threads and SMs
 Development Tools
Programming environments
• NVIDIA CUDA C
• CUDA Fortran Compiler
• AMD Stream
• BrookGPU / Brook+
• RapidMind Platform
• OpenCL (Apple) : INTEL
• DirectCompute – Supported by DirectX 10 and DirectX 11
• NVIDIA Nexus or NSight
 OpenCL
• Open Computing Language (OpenCL) is a framework for
writing programs that execute across heterogeneous
platforms consisting of central processing units (CPUs),
graphics processing units (GPUs), digital signal processors
(DSPs), field-programmable gate arrays (FPGAs) and other
processors or hardware accelerators.
• OpenCL specifies programming languages (based on C99 and
C++11) for programming these devices and application
programming interfaces (APIs) to control the platform and
execute programs on the compute devices.
• OpenCL provides a standard interface for parallel computing
using task- and data-based parallelism.
 Further Information
• Intel has now started pounding the
marketing drums on something long
predicted: integration of Intel’s graphics
onto the same die as its next generation
“Sandy Bridge” processor chip, due out in
mid-2011.
• Probably not coincidentally, mid-2011 is
when AMD’s Llano processor will see
daylight. It incorporates enough graphics-
related processing to be an apparently
decent DX11 GPU, the architecture hasn’t
been disclosed in detail.
• Animated movies (Wall-E Pixar)
– $180 million
• DreamWorks Animation Shrek 3 (2007)
– 1000 Linux Desktops, 3000 CPUs, 20
million render hours
• Blender - opensource rendering software
• Commercial 3D software
– Maya, V-Ray for Maya, 3ds Max, V-Ray
for 3ds Max, C4D, V-Ray for C4D,
Lightwave, XSI, V- Ray Standalone,
Modo, Cinema 4D and Brazil
• Commercially Cloud-based
Renderfarm
– Many still ‘traditional’ or non GPU
machines
– Extra Large (XL-CPU), 8 virtual cores (20
EC2 CUs), 7 GB RAM, 50 render nodes,
200 hours = $9742.42 (beta users =
$6843.03)
– Discount for >= 60 hours: 30 %
– $0.08 per core hour
• Built-your-own High Performance
Computer (HPC)
– 2 GFlops for $500
• Anyone interested in becoming GPU
specialist programmer? Or Computer
Graphics (CG) artist?