KB010: Research Computing Resource Catalog (4-Tier GPU/Compute Hierarchy)
Scope: All Compute & Storage Services
Audience: All Researchers
Last Updated: Feb 22, 2026 (Strategic Hierarchy Alignment)
☁️ Ursa Major (Google Cloud) - The Baseline
Best For: Generative AI, Compliance, Big Data.
- Strategic (Pool): Gemini API, Vertex AI, Secure Enclaves, Coldline Archive. (Subsidized, strict
e2-standard-2 limits apply).
- Commercial (Recharge): Custom VMs, Hot Storage. (Billable to Grant).
- Access: Google Cloud Console (SSO).
When advanced compute or GPU workloads exceed the subsidized Tier 1 limits of Ursa Major, researchers must follow the 4-Tier GPU Offload Hierarchy below:
The 4-Tier GPU & Compute Offload Hierarchy
1. 🖥️ UCR HPCC (Primary Offload)
The main and preferred destination to maximize ROI on local infrastructure.
Best For: Batch jobs, Simulation, GPU Training, MPI.
- Hardware: 6,000+ Cores, NVIDIA A100/H100/T4 GPUs, Infiniband.
- Storage: GPFS Parallel File System (Fast I/O).
- Cost: $1,000/year Flat Rate per Lab (Unlimited Access). Hardware Buy-in available.
- Access: SSH / OpenOnDemand.
A regional/national grid leveraging existing UCR hardware contributions.
Best For: Docker Containers, Web Services, Jupyter Notebooks that overflow HPCC.
- Hardware: Distributed GPU cluster across California.
- Storage: Ceph Object Storage.
- Cost: Free (Allocated via simple request).
- Access: Kubernetes (kubectl) / Portal.
3. 🇺🇸 NAIRR Pilot - National AI Resource (Tertiary)
For specialized, high-end AI allocations when local/regional capacity is insufficient.
Best For: Massive-scale AI, Exotic Hardware (Cerebras, SambaNova), hundreds of H100s.
- Cost: Free (Competitive Proposal - Startup or Research Allocation).
- Access: Application via nairrpilot.org.
4. 🇺🇸 NSF ACCESS (Quaternary)
Traditional national supercomputing grids reserved for massive-scale jobs.
Best For: Extreme-scale HPC requiring thousands of compute nodes.
- Cost: Free (Credit Exchange).
- Access: Application via access-ci.org.