What is the key feature of Graphical Processing Units (GPUs) that makes them well-suited to running Al applications?

The primary architectural advantage of Graphics Processing Units (GPUs) for AI applications is their massively parallel design. GPUs contain thousands of smaller, specialized cores, enabling them to perform a vast number of simple, repetitive calculations simultaneously. AI and deep learning workloads, which are fundamentally based on large-scale matrix and vector operations, are inherently parallelizable. This allows the computational load to be distributed across the many cores of a GPU, leading to a dramatic acceleration in processing speed compared to Central Processing Units (CPUs), which are optimized for sequential, low-latency execution of complex tasks.

B. CPUs are generally optimized for lower-latency memory access for sequential tasks, whereas GPUs prioritize high-bandwidth memory to service their many parallel cores simultaneously.
C. GPUs are specialized co-processors (accelerators) and cannot run every task; the CPU manages the operating system and general-purpose computations.
D. This is a misapplication of Moore's Law, which describes transistor density in integrated circuits generally, not a specific architectural feature of GPUs that makes them suitable for AI. ---

1. Stanford University

CS231n: Convolutional Neural Networks for Visual Recognition. In the "Hardware and Software" section of the course notes

it is stated: "The reason GPUs are so useful for Deep Learning is that they are very efficient at matrix/vector computations. It turns out that this is a very similar problem to rendering graphics on a screen

and as a result of the huge video game market

GPU manufacturers have spent billions of dollars of research on making very fast

very parallel chips (GPUs) that are great at matrix multiplication. ... GPUs have a massively parallel architecture (many

many cores)."

Source: Stanford CS231n Course Notes

Module 1

"Hardware and Software".

2. NVIDIA

CUDA C++ Programming Guide. The official documentation for the CUDA parallel computing platform explains the GPU's architecture. In the introduction

it describes the GPU as a parallel processor: "The GPU is especially well-suited to address problems that can be expressed as data-parallel computations... because it is designed as a parallel

multithreaded

manycore processor with tremendous computational horsepower and very high memory bandwidth."

Source: NVIDIA CUDA C++ Programming Guide

Version 12.4

Chapter 1

Section 1.2

"The GPU Architecture".

3. MIT 6.S191: Introduction to Deep Learning. The introductory lecture materials explain the hardware basis for deep learning's success. They contrast the CPU

with a few powerful cores optimized for serial tasks

against the GPU

with thousands of simpler cores ideal for parallel tasks like the matrix multiplications that dominate neural network training.

Source: MIT 6.S191: Introduction to Deep Learning

Lecture 1: "Foundations of Deep Learning

" section on "Hardware for Deep Learning".