1. Computer Graphics Graphics Hardware
CO2409 Computer Graphics
Week 12

2. Lecture Contents Graphics Architecture Processing: Memory:
GPU vs CPU
Memory:
Video Memory vs System Memory
Hardware Comparisons
Motherboard Interface
Parallelism / Concurrency

3. Graphics Architecture
The basic graphics architecture for all modern PCs and game consoles is similar:
Two key parts:
Main System
Graphics Unit
Local RAM for each processor
Fast access
Interface between components slow
Compared to local access

4. Graphics Unit
The graphics unit describes the graphics processing components of a system
PC: usually a separate graphics card
Console: built-in custom components
The core of a graphics unit is the Graphics Processing Unit (GPU)
Occasionally called a Graphics Adapter
Equivalent of a CPU for graphics
PS4 GPU

5. Graphics Unit The graphics unit also contains:
Local RAM (a.k.a. video memory, VRAM, etc.)
Output sockets for monitor, TV etc.
Can be integrated into the motherboard
Consoles, some PC’s and laptops
Attached via an interface:
PCs use PCI Express sockets
Can connect several together (SLI or Crossfire) for more parallelism
NVidia GeForce GTX 980 Graphics Cards
[Four connected by SLI]

6. GPU vs CPU A GPU is a dedicated graphics processor
Much more parallel than a typical CPU
i.e. Many more cores (100s or 1000s compared to 4-8 on a CPU)
Much faster than CPU for graphics algorithms
Particularly vector/matrix operations
But worse at general operations, esp. conditional instructions
Driven a programmable pipeline (shaders)
Can now perform some general programming using the flexibility of this – GPGPU programming (General Purpose GPU)
Power of a GPU is measured by its:
Clock speed
Parallelism. E.g. number of simultaneous shader, texture and output operations

7. Graphics Memory Many graphics units have local memory
To store vertices, primitives and textures for the GPU
This memory is similar to standard CPU memory
Usually more closely coupled to the GPU
I.e. Bandwidth to the GPU will be greater
Graphics memory can be measured by:
Clock speed (how fast can it serve data)
Bandwidth to the GPU (clock speed * width of the bus between graphics memory and the GPU)
Some GPUs have no dedicated memory
Must share memory with the CPU
This can be a benefit or a penalty…

8. Graphics vs System Memory
GPUs can usually access system memory
i.e. CPU-local memory
Bandwidth is often lower
Usually better to have graphics data in GPU memory
System memory can be used as a back-up
But expect lag if relied on too heavily
Some architectures share CPU & GPU memory
Xbox 360 / One / PS4: entire console architecture designed around this sharing – performance is high
On-board video cards: Lack of GPU memory is a cost-cutting feature – memory access is slow

9. Graphics Unit - Specifications
The key specifications for a graphics unit are:
GPU clock speed (MHz ) – speed of processor
Amount of Memory (Mb)
Memory clock speed (MHz / GHz)
GPU<-> Memory Bandwidth (GB / s)
Speed of transfers between GPU and local memory, determined by memory clock speed and bus width
Other factors to take into account:
Pixel / Texture Fill Rate (Giga-Pixels / s)
Speed it can output pixels to screen or textures
Amount of parallelism in the pipeline
E.g. Number of shader threads/cores – how many vertices / pixels can be processed at the same time

10. Historical Comparison (Indicative)
Platform
PS2
XBox
Xbox 360
PS3
Xbox One
PS4
GeForce
GTX 680
GTX 980
CPU / GPU Clock (MHz)
294/147
733/233
3 x 3200 / 500
1+7x3200 / 500
8x1750 / 853
8x1600 / 800
- / 1058
- / 1216
CPU / GPU Memory (Mb)
32 / 4
64 (Shared)
512
(Shared)
256 / 256
8000 (Shared)
- / 2048
- / 4096
Memory Clock (MHz) ?
200
700
650
2133 (GDDR3)
5500 (GDDR5)
6000
7000
Memory Bus (GB/s)
3.2 / 48
6.4
22.4
20.8 / 22.4
68 / 102x2
176
192
224
Pixel / Texture Fill Rate (GPx/s)
? / 0.932
? / 4
(No AA)
12.8 / 40.9
25.6 / 57.6
128.8
144
Parallelism -
4 / 2 Shader
48 Shader
24 / 8 Shader
768 Cores
1152 Cores
1536 Cores
2048 Cores

11. GPU / System Interface
It is faster to keep graphics data local to the GPU
But the CPU and system memory still need to interface with the graphics unit:
To get data into the graphics memory
To issue instructions to the GPU (DrawPrimitive)
For dynamic geometry / textures
So the interface between the graphics unit and the rest of the system is critical
Games consoles have the graphics unit built into the motherboard and so are closely coupled
E.g. Xbox 360/One, & PS3/4 have fairly symmetrical performance between GPU, CPU & memory

12. PC Graphics Interfaces
Interface between system and GPU on PC:
PCI Express
Uses serial ‘lanes’ to transfer streams of data in parallel
Version 3 supports 1GB/s transfer rates (theoretical).
Version 4 specs are being finalised
Has superseded the earlier interfaces AGP & PCI
But still can be a bottleneck
(Console memory bus much faster)
Slow compared to local GPU memory
PCs must rely on more graphics memory

13. Parallelism / Concurrency
GPUs are parallel architectures
Processing many pixels / vertices at the same time
A graphics application is typically a concurrent system
Graphics rendered while main application does other processing
Concurrent = different tasks performed simultaneously
Parallel = the same task split up and performed simultaneously
Need to program carefully to get best performance:
Ensure both CPU and GPU working all the time
Neither should be waiting for the other to complete its current task
But watch out for problems with shared data
Implications about programming graphics applications
Games students will see this in the 3rd year

14. General Purpose GPU A GPU is a massively parallel processor
Especially for vector maths operations
So it can be used for certain non-graphics tasks
Physics simulation, video processing, weather forecasting, etc.
Anything with massive amounts of mathematical calculation
Called General Purpose GPU processing (GPGPU)
Several APIs for this:
CUDA (from NVidia)
Extension to C language
Up to 40 times faster than same code on CPU
OpenCL for ATI and NVidia platforms
Compute Shaders as part of DirectX11