Scaling AI Education Without Large-Scale Local HPC: Shared Models for Emerging Institutional Capacity
Xiaolei Huang
University of Memphis
xiaolei.huang@memphis.edu
CRA NAIRR Pilot Classroom Expansion Conference� Research-Emerging Institutions with Large Computing Programs�Orlando, FL | June 18-19, 2026
Sharif, M., Han, G., Liu, W., & Huang, X. Cultivating Multidisciplinary AI Workforce Development on iTiger GPU Cluster: Practices and Challenges. PEARC 2026.
Computing Paradox and Dilemma
Research-Emerging Universities
~300
Increasing Demand + Capital Shortfall
$10B - $20B
Practical Capacity Issues in HPC Adoption
Question: How do we make shared AI infrastructure accessible, usable, and sustainable?
Cost and access
Local GPU clusters are expensive to acquire, operate, secure, and refresh.
Instructor readiness
Faculty may want to teach AI but lack examples that run reliably on shared systems.
Student readiness
Many students have limited experience with command line tools, schedulers, containers, or debugging GPU jobs.
Lessons from iTiger, a GPU Cluster�Computing Adoption and Challenges�2023 - 2026
CNS-2318210
NSF-MRI
iTiger Cluster for Research & Education
>300 users
226 active users
203 internal; 19 external
181 students; 45 faculty
Computing Adoption Strategies
Research Initiatives
CI Services
Education Efforts
New User Registrations
New users clustered around in-class tutorials and course-related training activities.
Job Submissions
Job submissions show strong course-driven student demand and growing regional research use.
Student use
UofM students account for most submissions, with peaks aligned to course activity.
Faculty and external use
Faculty and external researchers show lower but steadily increasing submission volumes.
Regional model
External users were a small share of accounts but contributed meaningful job submissions.
Job Outcomes
Completed jobs increased, but failed jobs also rose during intensive training and project periods. Temporary spikes in failures reflected experimentation, debugging, and workflow learning.
HPC Adoption in AI+X
Implication for NAIRR: Shared AI education capacity should support both computing and non-computing disciplines.
Takeaways for Teaching Support
Resource layer
What instructors need
What shared programs can provide
Compute
Reliable access for assignments
Small allocations, class queues, starter accounts
Software
Stable environments
Containers, notebooks, tested model recipes
Curriculum
Plug-in AI modules
Assignments, rubrics, domain datasets, socio-technical prompts
Support
Fast help during course peaks
Office hours, FAQs, instructor Slack/Discord, escalation paths
Assessment
Evidence of capacity growth
Activation, job outcomes, project artifacts, confidence surveys
Takeaways for Broadening HPC Adoption
Shared AI infrastructure per regional hub is necessary to motivate emerging institutions full participants in AI education and research.
Shared compute
Course integration
Human support
Regional community
Takeaways for Emerging Institutional Capacity
Takeaways for Sustainable Capacity
National Resource + regional Resource Hubs + local support = Sustainable AI education capacity.
National shared resources
Large-scale models, compute allocations resources (NAIRR)
Regional hubs
Local/regional support, community, and sharing network
Institutional
implementation
Research, instructors, courses, capstone projects
Thanks!
xiaolei.huang@memphis.edu
Acknowledgement: The presenter wants to thank the NSF and CRA for generous support!
itiger-cluster.github.io
Sharif, M., Han, G., Liu, W., & Huang, X. Cultivating Multidisciplinary AI Workforce Development on iTiger GPU Cluster: Practices and Challenges. PEARC 2026.