Higher Education and Research HPC Hub
Campus research computing sits between instructional labs and national lab–scale systems: shared GPU partitions for classes, discipline-specific HPC cells for funded projects, and storage tiers that absorb instrument ingest without starving interactive jobs. New Tech Solutions (NTS / RackMountNTS) designs CPU and GPU nodes, InfiniBand or RoCE fabrics, and scratch-to-archive storage for university and research-institute programs—then quotes the same BOM on cooperative and state vehicles when procurement requires it.
Why campus HPC and AI clusters fail without a coherent BOM
Research groups often grow by adding opportunistic nodes—different GPU SKUs, mismatched NIC firmware, and scratch that was never sized for checkpoint storms. Schedulers (Slurm, PBS, Kubernetes GPU operators) then spend more time on queue fairness and driver drift than on science throughput. A durable campus design starts with fair-share GPU pools, CPU MPI or analysis farms that match the dominant codes, a fabric plan that scales leaf-to-spine without forklift upgrades, and NVMe scratch placed next to the jobs that write it. NTS treats those layers as one engineer-validated configuration rather than a shopping list of unrelated SKUs.
Program patterns we support
- Shared campus GPU pools for teaching, workshops, and multi-PI research with partition quotas and accounting hooks
- Discipline-specific HPC cells (simulation, imaging, NLP, climate) with InfiniBand or RoCE and software-aligned CPU/GPU mixes
- Parallel scratch and tiered capacity near compute—NVMe for active jobs, capacity tiers for project and archive retention
- Pilot nodes that expand into multi-rack programs without rewriting imaging, BMC, or spare matrices
- Dual-vendor NVIDIA and AMD Instinct options when export control, software stacks, or pricing favor diversity
Scheduler-friendly nodes, fabric, and storage
NTS sizes hosts for the queue you actually run: dense GPU nodes for training and fine-tuning, dual-socket CPU nodes for classical HPC and ETL, and storage-adjacent compute when Ceph, Lustre-adjacent, or object gateways sit in the same cell. Fabric choices track message size and all-reduce behavior—HDR/NDR InfiniBand for tightly coupled jobs, RoCE where Ethernet operations standards dominate. Scratch capacity is derived from instrument rates and checkpoint intervals, not from a generic “add more SSDs” heuristic. Browse related platforms on HPC and AI clusters, GPU servers, and storage tiers.
Cooperative and SLED procurement for higher education
Many universities buy research IT through cooperatives (OMNIA, NASPO, TIPS, E&I, and state DIR or CMAS paths) rather than open market POs. NTS maps validated BOMs to those vehicles so research computing and contracting offices share one part list, lead-time narrative, and acceptance package. When federal research institutes buy on SEWP V, GSA MAS, or CIO-CS, the same technical baseline can be re-lined without redesigning power, fabric, or imaging. Start from the contracts hub or the public-sector procurement guide when vehicle selection is still open.
Typical research outcomes (anonymized)
Shared GPU partitions with fair-share scheduling and teaching overlays; NVMe scratch sized to cryo-EM or genomics ingest; dual-vendor Instinct and NVIDIA cells for software diversity; pilot racks that graduate to multi-rack AI teaching labs with unchanged image and spare kits. In each case NTS stages and burns systems in Fremont, CA, so campus data-center teams receive serial-tracked baselines instead of untested freight.
Teaching overlays and multi-PI fairness
Campus clusters rarely serve a single grant. Instructional partitions need predictable GPU hours during the semester; research partitions need longer walltimes and larger memory footprints. NTS helps research computing teams express those overlays in the BOM—node classes, MIG or time-sliced GPU options where appropriate, and accounting-friendly serial inventories—so ops can enforce policy without buying a second cluster. When a college pilots a generative-AI teaching lab, we size a starter cell that keeps the same image, fabric MTU, and spare kit the central cluster already trusts.
Facility and operations checklist
Before PO, confirm rack U, PDU phase balance, aisle containment, and whether liquid cooling is on the roadmap. NTS documents kW envelopes and airflow assumptions with the quote so facilities and research computing sign off together. After delivery, firmware baselines and BMC settings travel with the shipment so the first expansion rack does not invent a second standard. For remote evidence before commit, ask about POC or staged soak patterns that mirror your dominant codes.
How to engage NTS on a research cluster
Bring workload notes (codes, GPU vs CPU mix, concurrent users), facility limits (kW, cooling, rack depth), storage retention, and your preferred contract vehicle. NTS returns a right-sized BOM, fabric and scratch assumptions, staging scope, and a quote package contracting can review. If you already hold a draft part list, send it with power and delivery constraints—we flag PCIe budget, PDU draw, and fabric oversubscription before the PO locks the wrong assumptions.
Explore cluster solutions, data center and HPC, and AI infrastructure, or plan a research cluster with NTS.
