course: Fundamentals of GPU Programming

number:
141374
teaching methods:
lecture with tutorials
media:
overhead transparencies
responsible person:
Prof. Dr. Ralf Peter Brinkmann
lecturer:
Dr. Denis Eremin (ETIT)
language:
english
HWS:
3
CP:
4
offered in:
winter term

dates in winter term

  • start: Thursday the 12.10.2017
  • lecture Thursdays: from 14:15 to 15.45 o'clock in ID 03/471
  • tutorial Thursdays: from 16:15 to 17.00 o'clock in ID 03/471

Exam

Oral

Date according to prior agreement with lecturer.

Duration: 45min
Exam registration: FlexNow

goals

The students know how to program on graphics processing units (GPUs)

content

Upon a certain point in time around 2003, the computational performance has been increasing not at the expense of boosting the processor tact frequency, but rather through increase of the number of computational cores allocated on the processor chip. Graphics processing units (GPUs) are champions of this computer hardware evolution, boasting up to thousand single core units. As a result, GPUs arguably demonstrate far better raw performance of the arithmetical units and the memory system compared to his older "brother", central processing units (CPUs). This is due to a conceptually different architecture of GPUs compared to that of CPUs and more innovative memory system, some parts of which can be manually controlled through a corresponding API. While designed originally for video processing tasks, the enormous computational power of modern GPUs is commonly used to assist to CPUs or to take main part in solving a large variety of computational problems featuring (massively) parallelizable sections, bringing the teraflops scale high performance computing powers to laptop/desktop computers. The present course shows how CUDA C (extension of the C language designed for GPU programming) and the corresponding (very flexible!) CUDA runtime API framework can be used to accelerate execution of some typical programming patterns by a factor of 10 or more compared to that of CPU. Starting from the CUDA Programming Model, one will proceed to the CUDA Execution Model and go on further to consideration of fundamental conceptual, software and hardware issues helping to understand how GPUs work. Case studies of several problems involving massively parallel algorithms implemented on GPUs will be also elaborated on. The theoretical knowledge given during the lectures will be reinforced by a large number of practical examples that students will be able to work on at home.

requirements

none

recommended knowledge

C (programming language)