Clusters built by the people who racked them
CPU and GPU cluster design and build for engineering, geoscience and research workloads — interconnect, scheduling and storage throughput engineered together, because a cluster is only as fast as its slowest layer.
HPC, end to end
Cluster architecture
Node count, CPU and GPU selection, memory per core and topology designed against the actual solver or model, not against a benchmark nobody runs.
Interconnect
High-density low-latency fabric, cabled and validated. On a parallel workload the interconnect decides your scaling curve, and it is the layer most often under-specified.
High-throughput storage
Parallel or high-throughput storage sized to keep the nodes fed. A cluster waiting on IO is expensive idle silicon.
Scheduling
Job scheduler, queues, fair-share policy and resource limits configured so the cluster serves the whole team rather than whoever submits first.
Power, cooling and rack
Rack power density, containment and cooling coordinated with the facility. High-density HPC fails at the facility layer more often than at the compute layer.
Hardening and operations
Hypervisor and OS hardened to the standard your sector requires, then monitored and supported around the clock.
Clusters we have delivered
Client names stay private. The builds do not.
- Geoscience HPC cluster, Tier-III data centre, Egypt. Architected the solution, racked and stacked every node, structured cabling and cluster interconnect, high-performance storage, routing and firewalls, then hypervisor operations, monitoring and 24/7 support. From bare racks to a production seismic processing cluster.
- CPU and GPU clusters, Saudi Arabia. Delivered on site in the Kingdom by Stark engineering teams: high-density interconnect cabling, rack power and cooling coordination, cluster networking, high-throughput storage, and a hypervisor stack hardened to SACS-002 cyber security requirements, with ongoing operational support.
- National data centre, Egypt. Full physical build under strict security and confidentiality requirements, including compute, storage and virtualisation.
Architect, build, tune, operate
1. Architect
Workload profiled first. Whether you are memory-bound, IO-bound or interconnect-bound changes the entire design and the budget.
2. Build
Racking, power, cooling, cabling and interconnect validation. Our engineers are on site with cable in hand.
3. Tune
Scheduler policy, storage throughput and node configuration tuned against real jobs, with the achieved performance documented.
4. Operate
Monitoring, patching, capacity and 24/7 support against clear service levels.
Common questions
Can you deliver outside Egypt?
Yes. We have delivered CPU and GPU clusters on site in Saudi Arabia with our own engineering teams.
GPU or CPU?
It depends entirely on the code. Some solvers scale beautifully on GPU and some do not scale at all. We profile before recommending, because getting this wrong is an expensive mistake.
Can you work with our facility team on power and cooling?
Yes, and we insist on it. High-density racks fail at the facility layer more often than at the compute layer.
Do you support it afterwards?
Yes, including 24/7 monitoring and support against clear service levels. See managed IT.
Can you harden it to a client security standard?
Yes. We have hardened an HPC hypervisor stack to SACS-002, and our compliance team produces the evidence auditors ask for.
Tell us what the cluster has to run.
The application, the model size and the job profile tell us more than any node count. Send those and we will architect against them.
02 3537 5791