Blue Lion at LRZ: What a 30x Leap Means for Science
Industry leaders are preparing a new leadership-class system for the Leibniz Supercomputing Centre (LRZ) in Germany. Announced at ISC 2025 in Hamburg and scheduled for 2027, the Blue Lion system is based on the HPE Cray GX architecture and NVIDIA's next-generation Vera Rubin chips. LRZ expects around 30x the compute of its current flagship system, with fanless liquid cooling to cut energy use and reduce environmental impact.
In a recent conversation, Prof. Dieter Kranzlmueller, Chairman of the Board of Directors at LRZ, outlined why this configuration was selected, what workloads it will prioritize, and how HPC and AI will work side by side to accelerate research across disciplines.
Blue Lion at a glance
- Architecture: HPE Cray GX with NVIDIA's next-generation Vera Rubin chips
- Scale: Targeting ~30x the compute of LRZ's current flagship system
- Timeline: Announced at ISC 2025; planned operation in 2027
- Cooling: Fanless liquid cooling for higher efficiency and lower environmental impact
- Focus: Converged HPC + AI workloads for science and engineering
Why this system, and why now
Selection criteria centered on balanced performance: compute density, memory bandwidth, interconnect efficiency, software ecosystem maturity, and total cost of ownership. The team prioritized a platform that can run both traditional simulations and modern AI models without compromise.
The HPE Cray stack and NVIDIA's next-gen chips are expected to deliver strong node-level throughput and cluster-scale efficiency. Just as important, the software toolchain aims to give researchers a practical path from development to production at scale.
What Blue Lion will run
Expect large-scale simulations (climate, CFD, materials, cosmology), AI-assisted modeling (surrogate models, physics-informed learning), and hybrid workflows that mix simulation, data assimilation, and learning. The 30x step up opens doors to higher-resolution meshes, longer time horizons, and larger ensembles for uncertainty quantification.
On the AI side, teams can train foundation and domain-specific models, fine-tune efficiently, and run inference at scale close to their simulation data-cutting transfer overhead and improving turnaround time.
Cooling, energy, and environmental impact
Fanless liquid cooling reduces electrical overhead from fans and enables higher component density. Warm-water operation can improve overall efficiency and simplify data center design.
For projects with tight energy budgets or sustainability goals, this matters. More work per watt means more science within the same footprint and funding envelope.
What researchers can do now
- Profile current codes to identify GPU readiness, memory pressure, and communication hot spots.
- Adopt portable programming models and containers (e.g., MPI + GPU acceleration, well-defined images) to ease future migrations.
- Validate mixed precision where appropriate; quantify error bounds and speedups.
- Plan I/O carefully: checkpoint cadence, parallel file I/O, and data reduction strategies.
- Set up reproducibility: pinned environments, versioned datasets, and automated pipelines for experiments.
- Prepare hybrid workflows that link simulation, AI training, and inference in one loop.
LRZ's direction for the decade
According to Prof. Kranzlmueller, LRZ is doubling down on scientific impact, reliability, and sustainable operations. Expect continued investment in early-access programs, co-design with user communities, and support for software readiness.
The goal: shorten time-to-discovery by making HPC and AI work together-cleanly, at scale, and with clear paths from prototype to production.
Further reading
Upskilling for AI + HPC workflows
If your team is preparing to integrate AI into simulation pipelines, explore practical training paths: AI courses by job role.
Your membership also unlocks:
