GPU Options for the Project

Depending on the size of your training set and the speed of your CPU, you might be able to train your neural network on your local CPU. Training could take anywhere from 15 minutes to several hours if you train for many epochs.

A faster alternative is to train on a GPU.

Your Options

In the Deep Learning projects, you have a few potential options on how you can use GPU to speed up your network training.

Project Workspace

Udacity provides a workspace built into the classroom itself that supports GPU. You can toggle the usage of this GPU on and off to avoid using up the GPU hours (50) provided as part of the course. While this is the simplest easy option to use due to no set-up or charges, there can be potential latency issues (primarily related to the Behavioral Cloning project) you may need to consider. The instructions for using the workspace can be found later in this lesson.

AWS or other Cloud Compute

it’s easy (although not free) to access a GPU-enabled server (also known as an "instance") through Amazon Web Services (AWS). There are also other cloud compute services available for you to use GPUs with.

On the next page, we'll walk you through the steps of setting up an AWS account and launching your first instance.

Local GPU

It's possible to purchase your own NVIDIA GPU, or you may have one built into your machine already. If so, you can use the Term 1 Starter Kit along with your own local GPU to run on your own computer. The instructions for using a local set-up can be found later in this lesson.