EuroHPC JU has contracted Bull to deliver LUMI-AI, a planned supercomputer for the CSC – IT Center for Science data centre in Kajaani, Finland. According to AMD, the system will use Instinct MI430X GPUs and 6th Gen AMD EPYC processors with 256 cores, with deployment scheduled for the second half of 2027.

AMD says LUMI-AI is expected to provide 10 times the AI capacity and nearly twice the high-performance-computing capability of the current LUMI system. Those are company-provided projections: the announcement does not specify the benchmarks, workloads or measurement method behind the comparisons.

Contents

What is being built

LUMI-AI is intended to become the computing backbone of the LUMI AI Factory, a six-country consortium comprising Finland, Czechia, Denmark, Estonia, Norway and Poland. The system will be owned by EuroHPC JU and hosted at CSC’s new data centre in the Renforsin Ranta business park in Kajaani.

The procurement contract is valued at €387.8 million. AMD says that amount includes acquisition, delivery, installation and maintenance, so it should not be interpreted as the price of the processors and accelerators alone. Funding is described as being shared equally by EuroHPC JU and the LUMI AI Factory consortium.

The hardware plan combines two types of processors:

  • AMD Instinct MI430X GPUs: accelerator processors intended to handle highly parallel workloads such as AI training, inference and parts of scientific computing.
  • 6th Gen AMD EPYC 256-core processors: general-purpose CPUs that can coordinate workloads, process data and run system software alongside the accelerators.

The announcement does not give the planned number of GPUs or CPUs. It also does not specify system memory, networking hardware, storage capacity, peak performance or the software stack.

How the system is intended to work

High-performance computing, or HPC, uses large numbers of processing units and high-throughput data movement to run demanding scientific and engineering workloads. AI training and inference can use GPUs, while CPUs handle orchestration, preprocessing and other general-purpose system tasks.

LUMI-AI is planned as a multi-tenant system. In this context, multi-tenancy means that infrastructure can be allocated among different users or workloads rather than being dedicated to a single organisation. The source does not describe the system’s specific scheduling, isolation or resource-allocation mechanisms.

The project also plans to provide additional compute and storage resources with API-based access. This is intended to support software-mediated use of the infrastructure by different research and development groups.

The system is intended to support:

  • AI model training and inference;
  • scientific simulation;
  • academic research;
  • industrial AI development;
  • researchers, startups and industry users.

LUMI-AI is also planned to integrate with the LUMI-IQ quantum-computing platform. The supplied announcement does not explain how that integration will work or which workloads will use both systems.

Why the project matters

The project is more than a commercial server deployment. It is planned as shared public research infrastructure serving several European countries through the LUMI AI Factory.

If delivered as planned, LUMI-AI would expand access to large-scale AI acceleration while also supporting conventional HPC workloads. That combination matters because scientific users may need both AI methods and high-precision simulations, often alongside substantial data-processing and storage requirements.

The planned facility also links computing capacity with energy and heat management. AMD says the system will use 100% renewable energy and liquid cooling, with excess heat captured for reuse in Kajaani’s district-heating network.

Liquid cooling can transfer heat more directly than air cooling, and a suitable facility design can make waste-heat recovery practical. However, the actual energy performance and amount of reusable heat will depend on the complete data-centre design, workload and operating conditions. No operational measurements are available because deployment is scheduled for 2027.

What to watch before deployment

Deployment is planned for the second half of 2027. Before the system becomes available, important details to watch include the final accelerator and CPU counts, interconnect, storage architecture, software environment and access rules.

After installation, independent measurements will be needed to assess AI performance, HPC performance, energy efficiency and the practical recovery of heat for district heating. It will also be important to see how access is divided among academic researchers, startups and industry, and how LUMI-AI’s connection to LUMI-IQ is implemented.

Sources