DDNDDN
Opportunities

Asterion Research

ValidateActiveEXAScalerAI400X3MEMEAOn-premises
HPC checkpoint & data-access acceleration
£850.0K
71
Qualification
0
Solution Fit
0
Discovery
0
Value Conf.
90
Tech Confidence
Industry
Higher Education & Research
Country / Region
United Kingdom / EMEA
Close Date
—
Sales Owner
—
SE Owner
—
Probability (SE judgement)
55%
Quote Status
—
Contract Value
£2.40M
Next Action
Schedule benchmark plan review with specialist
Owner: Daniel Okafor · Due: 2026-12-01
Business Challenge & Consequences

Research throughput limited by checkpoint restart and data-loading delays on large simulation runs.

Fewer simulation iterations per quarter; longer time-to-result for competitive research programmes.

Desired Business Outcomes

Reduce checkpoint time and data-loading delays; increase productive GPU use and simulation throughput.

Compelling Event

New GPU cluster procurement requires matched data platform to hit ROI.

Workloads & Solution Direction
HPC and scientific simulationAI training and fine-tuning
Primary Direction
parallel-file
Application / Framework
—
Current Environment
Compute / GPU
8× H100 nodes under Slurm
GPU / Node Counts
64× H100
Storage / Data Platform
Competitor parallel file system, undersized
Network
200Gb InfiniBand
Workload Characteristics
Data Capacity & Growth
Hot working set ~60 TB today, projected to grow to ~120 TB within 24 months as simulation outputs and training datasets expand. Illustrative estimate (fictional).
Active Working Set
60 TB hot
Current Performance
10h per training run, 45 min checkpoint
Target Performance
8h per training run, 15 min checkpoint
Access Patterns
Sequential large-file, metadata-heavy at scale
Read/Write, Concurrency
Mixed read-dominant workload during training data loading (~80% read); high metadata intensity (stat/open) from many small files; up to 64-way concurrent GPU jobs under Slurm. Illustrative (fictional).
Required Protocols
POSIX, Lustre, MPI
Checkpoint / Restart
Checkpoints every ~2 hours during training; current ~45 min at 8-node scale; target < 15 min (acceptance < 12 min at 16-node; < 20 min at 64-GPU scale). Fictional baseline.
Availability / Recovery
Tier-1 research storage; N+1 appliance resilience; Lustre RAID; target RPO <= 30 min, RTO <= 4 h for the data platform. Illustrative target (fictional).
Security / Residency
On-premises deployment; research data must remain in UK; single-tenant research cluster (no multi-tenant isolation requirement). Fictional.
Known Technical Blockers
Metadata contention at high concurrency
Success Criteria (agreed)
Checkpoint time < 20 min at 64-GPU scale; sustained throughput target met
Competition & Alternatives
Competitors
Incumbent parallel file system vendor
Alternatives
Buy more GPUs, do nothing