Gpu and cuda general Purpose computing on

GPU and CUDA

• General Purpose computing on

• Graphics Processing Units

Introduction

• GPGPU is a technique for using GPUs to do work that is traditionally handled by the CPU

• Why?

• Programmability
• Precision
• Performance

Motivation?

• Computational Power!

• GPUs are FAST!
• CPUs follow Moore's Law
• GPUs speed increases faster than Moore's Law

GPUs Getting Faster, Fast!

• Specialized nature of the GPUs

• Transistors not cache

• Economics

• Huge video game industry means more money for development

GPUs are Flexible and Precise

• Modern GPUs are programmable

• Programmable pixel and vertex engines
• High Level Language Support

• Modern GPUs support high precision

• 32-bit floating-point throughout the pipeline
• High enough for many(not all) applications

Awesome Potential

• The performance and flexibility of GPUs makes them an attractive platform for general purpose computation.

• Clusters

• Cheaper high performance computers for schools

• Insane number of Gigaflops

CUDA

• Developed by NVIDIA

• An architecture the enables the use of standard programming languages on their graphics cards

• C for CUDA

• Third Party Wrappers

• Python
• Fortan
• Java
• MATLAB

CUDA

• Allows latest NVIDIA cards to have open architecture like a normal CPU

• But a GPU is a parallel “many-core” architecture

• Each core is capable of running thousands of threads

• Enables huge performance benefits

The Latest

• CUDA works with all NVIDIA GPUs from the G8x series onwards including GeForce, Quatro, and Tesla.

• Programs written for G8x series will work on all future GPUs

• Tesla chip designed specifically for CUDA programming

CUDA - Advantages

• Scattered Reads

• Shared Memory

• Faster downloads and readbacks

• Full support for integer and bitwise operations

Limitations

• Deviation from IEEE standards

• Bottleneck between CPU and GPU

• Threads should be running in groups of at least 32 for best performance

Process Flow

• 1. Copy data from main mem to GPU mem

• 2. CPU instructs the process to GPU

• 3. GPU execute parallel in each core

• 4. Copy the result from GPU mem to main mem

My GPU

• Way better than a CPU

• NVIDIA GeForce 8800 GTS

• Number of Multiprocessors:
• 12
• Number of Cores:
• 96
• Total Memory :
• 320 MB

Installation

• Started by installing Fedora 9

• Download the latest NVIDIA Driver, the SDK and the toolkit

• Install the Driver

• Install Toolkit

• Set paths

• Install the SDK

• If all goes well, your computer should be ready to compile and run CUDA programs

• Of course it doesn't

Setting the Paths

• Set the $PATH

• Export PATH=$PATH:/usr/local/cuda/bin
• All add this line to the ~/.bashrc file

• Set the $LD_LIBRARY

• Change the /etc/ls.so.conf file
• Add:
• /usr/local/cuda/lib

Problems

• First: needed to install binutils

• Second: install gcc

• Third: needed to install make

• Fourth: needed to install freeglut-devel

• Fifth: Wasn't seeing the nvcc binary file

• Everything was installed
• Path was right
• Architecture?
• Used arch command and got chip architecture (64-bit)
• But my OS architecture is 34-bit
• Everything I downloaded and installed was for 64-bit

More Problems

• Once I fixed the architecture problem I was able to run some of the sample programs.

• Some I would get errors on the ones with a graphical representation

• Had to install libXi-devel and the libXmu-devel

• Now Everything works perfectly

Using CUDA

• Compiler driver that simplifies the process of compiling C

• Simple and Familiar

• The SDK and comes with several sample test programs to run

• Stored in the projects folder of the SDK

Using CUDA (continued)

• Pick a desired test project

• Use the make command to compile the project

• results stored in bin/linux/release/

• Run the program

• ./program

• The result printed to the screen are just the time it too to run the program

Current Future Uses

• Accelerated Rendering of 3D glasses

• Real time cloth simulation

• Distributed Calculations

• Medical analysis simulations

• Physical Simulations

• Accelerated Encryption/Decryption and compression

Some Examples

• ssh to fang

• ssh to gw (ie gateway)

• ssh to root@192.168.166.218

• cd to the NVIDIA/projects directory

• Pick a project to make

• Run the project

My future Work

• For my masters project:

• Make a HPC with many Tesla GPUs
• Make a cluster of GPU computers
• Make a GPU server...?