From the Research Universities edition of September 26, 2026
KIT inaugurates HoreKa 2: >42k cores, 300 AI accelerators, ~33 PFLOPS; doubles sim/AI compute (€17M)
Karlsruhe Institute of Technology (KIT) / NHR@KIT Scientific Computing Center · Higher education research computing (Germany NHR) · Germany (Karlsruhe)
- Publisher
- KIT Presseinformation 081/2026
- Original publication
- September 25, 2026
- Source retrieved
- 2026-09-27
- Event date
- 2026-09-25
What happened
On 25 September 2026 KIT inaugurated HoreKa 2, a €17 million NHR-accessible supercomputer with more than 42,000 processor cores, 300 AI accelerators, and planned peak performance of about 33 PetaFLOPS. SCC director Martin Frank states the system doubles compute capacity for simulations and AI applications. Hot-water cooling and 100% renewable Campus Nord power are cited for efficiency; GPU partition supports AI workloads across earth system, engineering, materials, particle physics, and other disciplines. Phased rollout: CPU Onyx began earlier; GPU Jade (NVIDIA Blackwell) expands AI capacity.
Why it matters
Campus and national research-computing leaders evaluating AI/HPC upgrades need capacity, efficiency, and shared NHR-style access claims with clear limits—this is an inauguration with architecture specs, not measured queue SLOs or scientific discovery outcomes.
Evidence and measured results
KIT press release dated 25 September 2026 opened via HTTPS; core/accelerator/PFLOPS, €17M cost, doubling claim, and sustainability statements taken from the opened German page.
Limitations and uncertainty
Institutional self-report; PFLOPS is planned peak; 'doubling' is operator claim without published baseline table; no queue-wait or energy-efficiency ranking metrics in this page; German national context limits direct U.S. transfer.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-26; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Shared academic AI/HPC capacity is saturated while simulation and AI workloads compete for scarce accelerators.
- Stakeholders
- Research computing / SCC leadership
- VPR / research administration
- NHR allocation committees
- Faculty AI and simulation users
- Facilities / sustainability officers
- Discovery questions
- What is current CPU vs GPU utilization and wait by partition?
- Which AI workloads need Blackwell-class GPUs vs CPU-only?
- How is national NHR vs campus priority allocated?
- What efficiency or renewable-power constraints bind CapEx?
- Value hypothesis
- A staged €17M-class CPU+GPU upgrade with hot-water cooling can double stated sim/AI capacity for national academic users if allocation and training keep the system shared.
- Potential engagement
- 90-day pilot migration of 3–5 AI/simulation groups onto new GPU partition with utilization and wait baselines.
- Unsupported claims
- Guaranteed scientific discoveries
- Measured queue-wait SLOs (not in this release)
- U.S. campus cost parity
- Autonomous research-agent readiness
Pre-sales engineering
Role takeaway
- Fit
- National academic hybrid HPC with staged NVIDIA Blackwell GPU partition beside CPU Onyx under NHR@KIT ops.
- Architecture and integration
- Integrate Slurm/fair-share, AI modules, identity, storage, and GPU Jade alongside prior HoreKa; follow NHR migration docs for user cutover.
- Prerequisites
- Power and hot-water cooling validated
- Allocation policy published
- GPU software stack
- User migration training
- Baseline utilization metrics
- Constraints
- Planned peak PFLOPS not sustained application performance
- Phased GPU rollout
- German NHR access rules
- No agent eval harness specified
- Security
- Define multi-tenant isolation, data classification, and export-control paths before hosting regulated research on shared AI nodes.
- Proof of value
- Partition availability dates
- GPU/CPU utilization
- Pilot job success rates
- Independent efficiency measurement if claimed
Delivery
Role takeaway
- Work
- Finish GPU partition rollout, migrate pilots, publish service catalog and allocation SLOs, train consultants.
- Dependencies
- Facilities cooling/power
- Network/storage
- NHR identity/allocations
- PI pilot groups
- Ownership
- KIT SCC/NHR owns platform ops; PIs own science methods; ministry/NHR own national access policy.
- Skills and adoption
- Train users on GPU right-sizing, module stacks, and fair-share; train ops on Blackwell partition incidents.
- Governance checkpoints
- Allocation policy sign-off
- Security/data-class review
- Pilot review at 30/90 days
- Acceptance criteria
- GPU partition generally available
- Pilot groups productive
- Published capacity vs prior system baseline
- Risks
- Phased delays leave AI users on legacy GPUs
- Efficiency claims unvalidated
- Oversubscription recreates queues
- Transferability limits for U.S. campuses
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Hybrid CPU+GPU NHR cluster with hot-water cooling and staged GPU Blackwell partition; treat as shared national academic AI/HPC capacity beside prior HoreKa generation.
Governance
Who approves, reviews and stays accountable for outcomes?
NHR alliance access and KIT SCC operational ownership imply allocation and fair-share policies; inauguration does not publish AI-agent evaluation or research-data classification rules.
Security and privacy
What data, permissions and controls need testing?
Press release omits multi-tenant isolation, export controls, and research-data categories; plan security review before sensitive AI workloads.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
NHR-wide access plus KIT training expands accelerator availability beyond single-lab GPU islands; language and allocation literacy remain barriers for non-HPC specialists.
Procurement
What should contracts, pricing and exit terms secure?
Justify shared CapEx with capacity doubling and efficiency design, not brochure FLOPS alone; include cooling and renewable-power OpEx in TCO.
Operating model
Which teams own the service once it runs?
SCC as shared service for national academic users is the operating model that converts hardware into research access; inauguration alone does not define AI-agent governance.
Publication history
- 2026-09-26Research Universities · Issue 204 resources
Stable resource ID: kit-horeka2-nhr-ai-hpc-20260925