Planning HPC and AI data centres
AI applications, simulations and data-intensive analysis require infrastructure suited to their computing needs. We plan GPU and HPC systems together with networking, storage, power and cooling. For local AI, we also deliver the complete AI stack through to the digital employee.
HPC and AI workloads shape the architecture.
High performance computing (HPC) brings computing power, networking and storage together in a coordinated system. Simulations, scientific computing, AI training and inference have different requirements. We start with the applications and their measurable goals: throughput, response time, data volume, parallelisation and growth.
Sizing considers CPU and GPU performance, available RAM, GPU memory and software support. Representative benchmarks help identify bottlenecks. The largest hardware configuration is not automatically the most cost-effective solution.
Size networking and storage together
A cluster can only work as quickly as data can be supplied and results exchanged. We assess bandwidth and latency, connections between compute nodes, data access and checkpoints. Networking and storage are aligned with the actual load profile and planned scaling.
Distinguish training from inference
Training can tightly couple many accelerators over extended periods. Inference depends more on request volume, response time and available model memory. For both, we plan capacity, reserves and monitoring to suit the application.
Cooling for high power densities
High power densities place particular demands on heat removal, flow rates and temperature control. The choice of air cooling, direct-to-chip liquid cooling or another approach depends on the hardware, load profile and site.
We coordinate IT systems and cooling plant together. Redundancy, maintenance access, leak detection, potential waste heat use and operational measurement are all considered.
Liquid cooling needs a complete operating concept.
Direct-to-chip cooling removes heat directly from suitable components. The overall system also includes heat exchangers or coolant distribution units, pipe routes, hydraulic design and building-side heat rejection. Temperature and water-quality requirements depend on the technology used.
We consider leak detection, isolation, maintenance access and incident handling. Residual air-cooling loads from networking, storage and other components must still be removed. Heat reuse depends, among other factors, on usable temperature levels and a suitable heat customer.
Design power supply around real loads
Required rack power cannot be derived from a generic industry figure. The specific GPU configuration, utilisation, simultaneous loads and planned expansion are decisive. We use them to develop load profiles and size power, cooling and reserves.
We assess economics and energy efficiency using documented assumptions. PUE, water demand and usable waste heat are considered within clear measurement boundaries, allowing transparent comparison of alternatives.
Load changes and brief power peaks are checked against the technical specifications of the selected systems. UPS, distribution and controls must suit the planned operating conditions.
HPC in an existing data centre
Before expanding, we check power and cooling reserves, weights, delivery routes, space for additional equipment and available network connections. Constraints become visible before hardware procurement. Together with operations, we develop a phased implementation plan.
Operation, availability and security
Job scheduling, monitoring, maintenance windows and recovery are part of the plan. Access rights, network segmentation and separate responsibilities connect IT with facility operations. Availability is defined by the applications and the consequences of failure.
What our HPC planning delivers.
- Requirements profile: applications, performance targets, data volumes and expected growth.
- Infrastructure concept: compute nodes, networking, storage and expansion stages.
- Utilities concept: documented load assumptions for power, cooling and reserves.
- Integration planning: building interfaces, installation routes, maintenance and, where necessary, construction phases.
- Acceptance framework: agreed performance, functional and operational tests.
We compare investment and ongoing costs, including energy, cooling, maintenance and operation. PUE expresses data centre infrastructure energy use relative to IT energy; it does not measure productive computing output. That requires workload-specific metrics and clear measurement boundaries.
Our overarching Data centre design connects the HPC concept with the entire site infrastructure. For local applications, we also provide sovereign AI and agentic AI consulting.
From HPC infrastructure to practical AI
For sovereign AI, we connect hardware with local models, a knowledge base governed by access permissions and your team’s applications. The digital employee can search documents, create drafts and support approved workflows.
Setup, integration, training and support form part of the agreed scope. Monitoring, updates, backup and responsibilities are planned from the outset.
Frequently asked questions
Does every AI system need liquid cooling?
No. Suitable cooling depends on hardware, power density, manufacturer requirements and the site. Air-cooled systems may be appropriate; higher densities may require other cooling approaches.
How much electrical power does a GPU rack need?
A single blanket figure is not meaningful. We determine demand from the specific equipment, load profile, operating model and planned reserves.
Can an existing data centre accommodate HPC?
Individual areas can often be adapted. Whether this is practical depends on power supply, heat rejection, space, structural capacity and operating requirements. An assessment of the existing facility establishes the limits.
Technical background
Current reference architectures treat computing power, networking, storage and building services as one coordinated task. Further reading: NVIDIA on AI factory architecture. The specific design is based on your applications and site conditions.