CUDA Programming – CPU vs GPU
Contents
●
Limitations of Multi-core CPU
●
Introduction to Graphics Processing Unit (GPU) ● GPU
Accelerated Applications
●
CPU vs GPU
●
GPU Hardware Architecture
2 22/06/19
Multi-core CPUs
Intel Xeon E7- 3.3 GHz 24 cores @ 9700k 8-cores
8855 v4 Intel Xeon E7- 3.4 GHz @4.9 GHz
14 cores 4850 v4 Intel Xeon E7- Intel Xeon
@ 2.80 GHz 16 cores @ 8867 v4
E7- 8870 v4
Intel Xeon E7- 2.80 GHz 18 cores @
20 cores
8880 v4 Intel Xeon E7- 3.3 GHz
@ 3.0 GHz
22 cores @ 8890 v4 Intel Core i7-
And so on...
3 22/06/19
If multi-core is solution!
Why can’t we increase to thousands of cores?
Why only few tens?
4 22/06/19
If multi-core is solution!
Cost
Complexity
Size
Heat/Power
# of cores
Limited speed
22/06/19
Other solution!
Graphics Processing Unit
11 22/06/19
History of GPU
●
The term GPU has been used from 1980s
●
Popularised by NVIDIA in 1999 who marketed the
Geforece 256 as “The world’s first GPU”
●
Initially intended for graphics related computing - To
accelerate the gaming and animation performance
●
In 2007 NVIDIA launched Compute Unified Device
Architecture (CUDA) which enabled General Purpose
Computing.
●
Now it is referred as General Purpose GPU (GPGPU)
6/19
Graphics Processing Unit (GPU)
Accelerator
Not a Standalone
Throughput oriented
179
Some of the GPU application areas
Games & Movies
Molecular Science
Modeling CFD Applications
Artificial Intelligence
8 22/06/19 What is GPU
Acts as an accelerator/Co-processor
Control
Compute Intensive task
Acts as an accelerator/Co-processor
Heterogeneous Computing Architecture
Not an intelligent device
Contains thousands of cores over millions of threads
can be launched
Not a standalone device
Takes orders from the CPU
Cannot replace CPU by GPU
23 22/06/19
CPU vs GPU
A single core CPU
1. Powerful
2. Need a lot of power
3. Complex control hardware
4. Good performance
CPU vs GPU
Many-core GPU
1. Less powerful but lot many cores
2. Require less power
3. Simple control hardware
4. Good throughput
Latency Oriented CPUs!!!
Note: These are very Powerful cores
In terms of frequency, transistors,
IPC, branch prediction etc.
30 22/06/191 22/06/19
Throughput Oriented GPUs!!!
Note: These are very light weight cores
In terms of frequency, transistors, IPC, branch prediction etc
32 22/06/19
Feature CPU GPU
Number or 1, 2, 8 or Few Thousands
cores hundred
Intelligence More Less
Standalone Yes No
Intends Latency oriented Throughput oriented
Core clock Higher Eg. 2.3 GHz Lower Eg. 900 MHz
rates
Efficiency Sequential Parallel
Power More powerful cores Less powerful
Usage General purpose Special purpose
Role Processor Co-processor