Research computing has three simultaneous jobs: enable discovery, teach people safely and preserve enough evidence to reproduce results. Hyperion systems can start as a principal-investigator node and grow into a scheduled departmental facility while keeping a consistent server, software and support architecture.
Match the platform to the research pattern
AxiFoundry supports method development and GPU teaching; AxiKeystone serves memory-rich CPU workloads; AxiAnvil and AxiForge address large models and accelerated science; AxiVector supports two-node MPI; AxiLattice provides a shared CPU facility; AxiCrucible, AxiBastion and AxiTitan support multi-user AI and mixed services.
Buy against named research codes and datasets. A benchmark appendix is stronger evidence for a grant or capital case than a theoretical peak number.
Create a reproducible software catalogue
Offer versioned modules or Spack environments for compilers, MPI and libraries, plus Apptainer images for complete workflows. Publish supported combinations and deprecation dates. Let projects bring containers through a review path instead of giving every user administrator rights.
Capture job manifests and encourage repositories containing workflow definitions, parameters and environment recipes. Persistent identifiers for datasets and images make papers and student work easier to reconstruct.
Teaching without destabilising research
Use partitions, quotas or reservations for taught sessions. Provide small datasets and bounded jobs that demonstrate scheduling, parallelism and GPU profiling without monopolising the platform. Resettable container environments let students experiment while protecting the host.
Teach performance literacy: serial baseline, speed-up, efficiency, memory fit, I/O and energy. Students should learn why more cores or GPUs can make a badly partitioned workload slower.
Fair access and research governance
Define project sponsorship, user onboarding, storage quota, queue limits, priority and data classification. Publish the rules and make exceptions reviewable. Sensitive clinical, commercial or export-controlled research may require dedicated projects, network controls and audit evidence.
Scheduler accounting can support transparent usage reporting, grant attribution and future capacity proposals without treating utilisation alone as scientific value.
Data management and preservation
Separate high-speed scratch, active project storage and preservation. Link outputs to the inputs, code and environment that created them. Retain enough metadata to rerun important results even when raw temporary intermediates are deleted. Test restore paths and decide who owns data when researchers leave.
A staged facility roadmap
- Start with one workload-led node and a documented environment.
- Add identity, backup, monitoring and a lightweight allocation process.
- Introduce Slurm when users or queues compete.
- Add RDMA-connected nodes only after scaling tests demonstrate value.
- Formalise service levels, training, governance and lifecycle funding.
- Review scientific outcomes and unmet demand before each expansion.
Primary technical references
- Slurm Quick Start Administrator Guide
- Apptainer user guide
- Spack documentation
- MLflow Tracking documentation
- NVIDIA CUDA C++ Best Practices Guide
References are provided for software architecture and implementation planning. Validate the versions, licences, support matrix and regulated-use requirements applicable to the final deployment.