Reading Assignment I: A Closer Look at GPUs

This week reading assignment provides an overview of the initial hardware design concepts that have been at the base of GPU development. Despite the paper is almost ten years old and many advancements have been done in this field, still the main motivations for GPU architecture still remain the same.

The paper first introduces The Graphics Pipeline concept - more about this in third module of the course - together with its execution characteristics (parallel processing, large variation in pipeline workload, potential use of instructions streams, i.e. SIMD, ...), after which GPUs were designed.

The paper then emphasizes key architectural choices of GPUs, such as Multicore + SIMD Processing = Lots of ALUs and Hardware Multithreading = High ALU Utilization.

Finally, the paper focuses on the GPU memory system (few things are now different in modern GPU) and on task scheduling and control. Particular interesting is the Running a Fragment Shader on a GPU core insert to fully understand execution on GPUs.

You can read the paper here:

AcloserLookAtGPUs-ACM-comm-2008.pdf Ladda ner AcloserLookAtGPUs-ACM-comm-2008.pdf

I hope you enjoy the reading. Possibly, you can discuss things you didn't understand in the paper in the discussion section of this module.