CC* Compute - Region: NCShare Accelerating Impact - GPUs-as-a-Service (NCShare AI-GaaS)

NCShare AI-GaaS addresses two extremes on the demand spectrum for GPUs: first, GPUs are often unavailable or unaffordable to institutions with modest needs and budgets; and second, Artificial Intelligence research, including Large Language Model research, often requires GPUs at a scale even the largest research universities struggle to attain. By moving GPU sharing from the individual campus level to the state level the result is more efficient GPU utilization, greater aggregate computational capacity for solving research problems, improved economies of scale, and democratized access for smaller schools and under-resourced minority-serving institutions.

NCShare AI-GaaS creates a powerful cluster of InfiniBand-interconnected Nvidia GPUs. The horizontally scalable cluster can be virtualized using the vendor's Multi-Instance GPU (MIG) with the slurm scheduler or their vGPU service and Kubernetes. These enable allocation and dynamically-configured access at the sub- or multi-GPU level, in support of a wide range of research and education needs.

Operating atop North Carolina's Research and Education Network, the cluster leverages two prior NSF awards that created a state-wide Shared Science DMZ and a shared High Performance Compute cluster providing virtualized software stacks. Combining existing NCShare federated access control with Confidential Computing capabilities on the GPUs, NCShare AI-GaaS can deliver secure, authorized GPU access to colleges and universities throughout North Carolina.

While the project is designed to serve general (non-regulated) research projects, a reference implementation using encrypted overlay networks and SDN techniques may enable use for certain regulated research projects.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.