Introduction

This lesson will introduce the concept of accelerated computing hardware, including discrete Graphical Processing Units (GPU) and the new Accelerated Processing Units (APU), and discuss how they differ from each other and from the traditional Central Processing Unit (CPU).


After defining much of the terminology likely to be encountered, it will discuss the architecture underpinning these different approaches to accelerated computation. In particular, the difference between a unified shared memory and a more traditional discrete memory set-up will be detailed, along with the advantages that this can grant an APU.


This will be followed by a discussion on the advantages and potential drawbacks of these different architectures, which will help to inform decisions on the best paradigm to choose given a particular use-case.

Learning objectives

By the end of this lesson, you will be able to:

  • Categorise different CPU and GPU architecture configurations
  • Understand the architectural underpinnings of the AMD Instinct GPUs
  • Evaluate which architecture is most suitable and cost effective for your application workload.

Presentation

This presentation, delivered by Dr Bob Robey from AMD, presents an overview of APU's versus discrete GPUs. 


Presentation (PDF)

Q1:

Guest users do not have permission to interact with embedded questions.


Q2:

Guest users do not have permission to interact with embedded questions.


Q3:

Guest users do not have permission to interact with embedded questions.

Key Points

  • An APU combines a CPU and GPU into a single package with shared, unified memory, simplifying programming.
  • A discrete GPU is a separate processor with its own dedicated memory, connected via PCIe.
  • The key advantage of an APU is its unified physical and virtual memory, which eliminates complex data copying.
  • Use an APU for workloads with heavy data transfer between CPU and GPU, or for easier application porting.
  • Use a discrete GPU for heavy computational tasks that require maximum processing power on the device itself.