The following are common questions from researchers getting started with GPUs. If you can’t find the answer to your question below or on the GPUs for Machine Learning page, please reach out to the research computing experts at UW-IT via firstname.lastname@example.org.
There is a level of resource abstraction when working with GPUs remotely whether in HYAK, the cloud, or another managed system. Whereas a system physically in front of you has limitations based on what you have assembled, the remote options are “assembled” using commands. For example, you may know you have a desktop machine with 1 GPU that you have purchased and built in front of you but if you use HYAK or the cloud you will have to use the SLURM or cloud commands, respectively. These commands are required to virtually build a system of equivalent resources for you to use. Some researchers consider this abstraction “virtualization”, however, it is important to note that it is not.
Yes. It is common, for example, within the Computer Science and Engineering department at UW, for a research group to purchase dedicated GPU hardware, either internal or externally mounted. Typically this is a cost/computation based decision.
There are roughly four criteria to consider:
- Your compute tasks “fit” on the compute capacity you are paying for
- You have the skills necessary for the build
- You are ok with the risk associated with a purchased hardware resource failing
- You have evaluated the alternatives (Hyak , cloud) as too expensive
- Colab is a consumer application where the contract is between the User and Google, and is not covered by Privacy or Data Security contracts with the UW. JupyterHub for Teaching is a potential alternative for courses.
- UW-IT cannot provide support for Google Colaboratory. Research computing experts at UW-IT may be able to provide a helping hand for researchers using or determining if Colab could help with a researcher’s needs.
Read about the powerful uses of GPUs for research computing on GPUs for Machine Learning.
It depends on many factors. Send a note to email@example.com, with the subject line: “Help with GPUs” to explore the alternatives and potential pricing.
- Software licensing in some cases is as easy as a checkbox. In other more complex situations it becomes a process. To first order using a software license on a cloud-based Virtual Machine (VM) is equivalent to using that license on “owned” hardware. The cloud VM is connected to the Internet and this plus some authentication of ‘right to use’ is all that is required. We even have recourse to the Linux sudo command available for root-level actions if necessary.
- In the case of a License server behind a firewall there is more of an issue. What this means is that a machine or machines outside the firewall (e.g. in the cloud) will need some sort of sanctioned pathway to the license server. This might be the case for a university-wide license for MATLAB for example. The “sanctioned pathway” would be handled through a support desk ticket with the agency that manages the firewall; for example using firstname.lastname@example.org.
- There are other possible complexities to consider as well. For example will the licensed software work effectively with GPU acceleration? Is the computational task best addressed by means of a cluster of two or more VMs? In either case it is simplest to contact the software vendor and initiate a conversation.
- Throughout these (more “process” than “checkbox”) scenarios UW research computing staff are available to help walk you through to a successful resolution.
We suggest sending an email to email@example.com to initiate a consulting help-desk ticket. This has worked well because the research computing staff have resources to draw on by virtue of years of networking with the UW research community. We may be able to connect you with other teams who have built comparable solutions. For example there are many projects using GPUs for convolutional neural networks (CNNs) at UW. Beyond ‘birds of a feather’ networking we can also help advertise for talented students, look into third-party consulting, and search further afield across academia through our wider-scale projects such as the West Big Data Innovation Hub and the CloudBank project.