Lecture 19

Learning objectives

After this class, you should be able to:

1. Explain the position of the GPU as a part of the PC architecture.
2. Use the above knowledge to identify performance bottlenecks for applications.
3. Explain the purpose of the cudaHostAlloc call and how it can be used for zero copy, which has the potential to improve performance.

Reading assignment

1. UIUC Lecture 20.
2. Zero Copy slides on Blackboard.
3. Lec 19 example code.
4. Read the article on PCIe available at: arstechnica.com/features/2004/07/pcie.
5. Look up internet resources to learn about zero copy.

Exercises and review questions

• Exercises and review questions on current lecture's material
  1. Which of the following three factors is most likely to limit vector addition performance: (i) PCIe bandwidth, (ii) device DRAM bandwidth, or (iii) GPU computing speed?
  2. Implement zero copy for vector addition and compare with performance without it. Report the performance results on Blackboard.

• Preparation for the next lecture
  1. Review double bufferring.

---

Last modified: 26 Mar 2015