Skip to main content

AMD Instinct and EPYC Power Europe's Next LUMI-AI Supercomputer

·1210 words·6 mins
AMD LUMI-AI Instinct MI430X EPYC Supercomputer HPC AI ROCm EuroHPC
Table of Contents

AMD Instinct and EPYC Power Europe’s Next LUMI-AI Supercomputer

Europe is moving ahead with its next-generation AI and high-performance computing infrastructure.

The EuroHPC Joint Undertaking (EuroHPC JU) has signed a €387.8 million procurement contract with Bull to deploy LUMI-AI at the CSC – IT Center for Science data center in Kajaani, Finland. Deployment is scheduled to begin in the second half of 2027.

As the successor to the existing LUMI supercomputer, LUMI-AI is designed to deliver approximately 10× the AI capacity of the current system while nearly doubling its conventional HPC performance.

The platform will combine AMD Instinct MI430X accelerators with 6th Gen AMD EPYC processors, creating a heterogeneous infrastructure designed for both AI and traditional scientific computing.

🌍 LUMI-AI Becomes the Compute Core of Europe’s AI Factory
#

LUMI-AI will serve as the compute core of the LUMI AI Factory, a European initiative intended to make advanced AI and supercomputing resources more accessible to researchers, startups, and industry.

The LUMI AI Factory consortium includes:

  • Finland
  • Czech Republic
  • Denmark
  • Estonia
  • Norway
  • Poland

The total procurement contract is valued at €387.8 million, covering the system’s procurement, delivery, installation, and full lifecycle maintenance.

Funding is split equally between EuroHPC JU, through the Digital Europe Programme, and participating consortium members.

LUMI-AI is therefore more than a conventional supercomputer replacement. It is being positioned as a shared European infrastructure layer for AI development, scientific discovery, and industrial innovation.

🚀 10× More AI Capacity, Nearly 2× HPC Performance
#

The headline performance target is a major increase in AI capability.

Compared with the current LUMI system, the new platform is expected to provide:

Metric LUMI-AI Target
AI capacity ~10× LUMI
HPC performance Nearly 2× LUMI
Deployment H2 2027
Contract value €387.8 million
Location Kajaani, Finland

The system will target a broad range of workloads rather than optimizing exclusively for AI.

Expected applications include:

  • Large-scale AI training
  • AI inference
  • Fine-tuning and scientific AI
  • Industrial AI development
  • Academic research
  • Scientific simulations
  • Data-intensive modeling
  • High-performance computing

This workload diversity is central to LUMI-AI’s architecture.

🧠 AMD Instinct MI430X Brings HBM4 to HPC and AI
#

At the accelerator level, LUMI-AI will use AMD Instinct MI430X GPUs, AMD’s accelerator platform for demanding HPC and AI workloads.

The MI430X is based on AMD’s CDNA 5 architecture and is specified with:

MI430X Specification Value
Architecture CDNA 5
HBM 432 GB HBM4
Peak FP64 288 TFLOPS
Target workloads AI + HPC
Software ecosystem AMD ROCm

The combination of 432 GB of HBM4 and high FP64 throughput is particularly relevant to scientific computing.

Traditional AI accelerators often emphasize lower-precision matrix operations for neural-network workloads. Scientific simulations, however, can require much higher numerical precision.

The MI430X is therefore intended to support both sides of the workload spectrum:

                    LUMI-AI
             ┌─────────┴─────────┐
             │                   │
          AI workloads        HPC workloads
             │                   │
      Training / Inference   Simulation
      Fine-tuning            Modeling
      Data analysis          Scientific computing
             │                   │
             └─────────┬─────────┘
                 MI430X + HBM4

This convergence is one of the defining characteristics of the new system.

⚙️ 6th Gen EPYC Venice Provides 256-Core CPU Compute
#

The accelerator layer will be paired with 6th Gen AMD EPYC processors, codenamed Venice and based on the Zen 6 architecture.

Flagship configurations can reach 256 CPU cores, providing the general-purpose processing capacity needed to coordinate accelerators, handle operating-system workloads, process data, and execute CPU-oriented HPC tasks.

A simplified division of responsibilities looks like this:

6th Gen EPYC CPUs
        ├── General-purpose computation
        ├── Data preparation
        ├── System orchestration
        └── CPU-side HPC workloads
            AMD Instinct MI430X
        ┌───────────┴───────────┐
        │                       │
   AI acceleration         HPC acceleration
        │                       │
 Training / inference      Scientific simulation

The combination gives LUMI-AI a heterogeneous compute platform instead of treating CPUs and GPUs as independent resources.

💧 BullSequana XH3500 Handles the Thermal Challenge
#

A system of this scale and density creates a substantial cooling challenge.

LUMI-AI will use Bull’s BullSequana XH3500 liquid-cooling architecture, allowing the system to remove heat directly from high-density computing hardware.

Liquid cooling is increasingly important for modern AI and HPC systems because accelerator and CPU power densities continue to rise.

The LUMI-AI deployment will take place at the existing CSC data center in Kajaani, Finland, extending Europe’s supercomputing infrastructure at a location already associated with large-scale scientific computing.

🔬 AI and HPC Converge in One Infrastructure
#

One of LUMI-AI’s most important design goals is workload convergence.

Instead of building separate infrastructures for AI training and traditional scientific simulation, the system is intended to support both within the same computing environment.

This matters because AI is increasingly becoming part of scientific workflows.

A modern research pipeline may look like:

Scientific data
Simulation ───────────────┐
      │                   │
      ▼                   │
Large-scale dataset       │
      │                   │
      ▼                   │
AI model training         │
      │                   │
      ▼                   │
Prediction / analysis ◄───┘
New scientific insight

For example, AI models can analyze simulation results, accelerate portions of scientific workloads, identify patterns in experimental data, or provide surrogate models for computationally expensive simulations.

LUMI-AI is designed around this increasingly blurred boundary between HPC computation, AI training, inference, and data analysis.

🌐 Building European AI Sovereignty
#

The project also has a strategic dimension.

Europe’s demand for AI compute is growing rapidly, but access to large-scale accelerator infrastructure remains concentrated among a relatively small number of global cloud and technology providers.

By investing in publicly accessible supercomputing infrastructure, EuroHPC and its member states can provide researchers and European companies with access to large-scale AI resources within the European research ecosystem.

LUMI-AI is particularly significant because it combines:

  • Large-scale AI acceleration
  • High-precision scientific computing
  • High-capacity accelerator memory
  • Open software through AMD ROCm
  • European research infrastructure
  • Industrial AI access

This makes the system part of a broader effort to build an AI ecosystem around European supercomputing resources.

📅 Deployment Begins in the Second Half of 2027
#

Under the current procurement plan, deployment of LUMI-AI is scheduled to begin in the second half of 2027.

Once deployed, access will be opened progressively to eligible European researchers and industry users.

The transition from LUMI to LUMI-AI therefore represents more than a generational performance upgrade. It marks a shift toward supercomputers that are designed from the beginning around the convergence of AI and traditional HPC.

With AMD’s MI430X accelerators supplying high-capacity HBM4 and AI/HPC compute, alongside 256-core Zen 6 EPYC processors and liquid-cooled Bull infrastructure, LUMI-AI is positioned to become a major European platform for scientific AI and large-scale computing.

🔮 LUMI-AI Points Toward the Future of Supercomputing
#

The architecture reflects a broader change across the supercomputing industry.

The traditional distinction between a supercomputer for scientific simulation and a data center for AI is becoming increasingly difficult to maintain. AI is now embedded in scientific research, while HPC workloads increasingly rely on machine learning and data-driven methods.

LUMI-AI addresses that convergence directly:

          Traditional HPC
        ┌───────────────┐
        │    LUMI-AI    │
        │               │
        │  EPYC + MI430X│
        │     + HBM4    │
        └───────────────┘
             AI / ML

If its deployment targets are achieved, LUMI-AI will give Europe a substantially larger platform for both AI and scientific computing while extending the capabilities of the existing LUMI ecosystem.

The key idea is not simply more GPU performance. It is the integration of AI acceleration, high-precision HPC, large-capacity memory, and general-purpose CPU computing into a single European-scale infrastructure.

Related

AMD ROCm 7.14 Adds Early Support for Ryzen AI MAX PRO 400
·743 words·4 mins
AMD ROCm Ryzen AI Gorgon Halo AI HPC PyTorch JAX GPU Computing
AMD to Invest £2 Billion in UK AI Innovation and Research
·665 words·4 mins
AMD Artificial Intelligence United Kingdom Data Center HPC AI Infrastructure Research EPYC Instinct GPU Technology Investment
AMD Instinct MI430X Targets HPC With 200 TFLOPs FP64
·1008 words·5 mins
AMD Instinct MI430X HPC FP64 Supercomputing HBM4 Exascale Computing GPU Architecture Scientific Computing AI Accelerators