Quick Answer: The Daedalus supercomputer Greece facility in Lavrio delivers 85.69 petaflops of verified performance, ranking 31st globally on the TOP500 list. Built on HPE systems equipped with NVIDIA GH200 Grace Hopper Superchips, this €58.9 million sovereign facility provides high-performance computing for climate modeling, emergency civil protection, localized healthcare diagnostics, and foundational Greek-language AI model training.
Greece spent decades outsourcing its most complex computational workloads to computing centers in Germany, France, and Italy. That reality shifted with the inauguration of the Daedalus supercomputer Greece deployment at the Lavrio Technological and Cultural Park. Delivering a benchmarked 85.69 petaflops of double-precision performance, this system converts a century-old industrial power station into a regional high-performance computing node. For systems architects, scientific researchers, and public infrastructure planners, this machine fundamentally reshapes how the Mediterranean region handles predictive fluid dynamics, localized climate modeling, and regional language model training.
Inside the Architecture: HPE, Grace Hopper, and Direct Liquid Cooling
Building a modern high-performance computing (HPC) system at this density requires solving severe thermal and bandwidth bottlenecks. Hewlett Packard Enterprise (HPE) architected the Daedalus system using the NVIDIA GH200 Grace Hopper Superchip platform, integrating CPU and GPU architectures onto a single module.
Traditional HPC clusters suffer heavily from PCIe bus limitations. When an accelerator finishes a calculation, shuffling large arrays back to host memory over PCIe Gen 5 creates massive latency spikes. The Grace Hopper architecture eliminates this bottleneck using the NVIDIA NVLink-C2C interface, which supplies 900 GB/s of bi-directional bandwidth between the 72-core ARM Neoverse V2 Grace CPU and the Hopper GPU. This coherent memory space allows the GPU to directly access the system's LPDDR5X memory without copying data back and forth through standard bus drivers.
+-------------------------------------------------------------------+
| NVIDIA GH200 SUPERCHIP MODULE |
| |
| +--------------------+ NVLink-C2C +----------------------+ |
| | ARM Grace CPU |<-------------->| Hopper GPU | |
| | (72 Neoverse V2) | 900 GB/s | (HBM3 / HBM3e) | |
| +--------------------+ Coherent Bus +----------------------+ |
| | | |
| 512GB LPDDR5X System RAM 96GB/144GB HBM3 VRAM |
+-------------------------------------------------------------------+
Thermodynamics pose an equally steep challenge. Air cooling at this density is impossible; standard server fans would consume more power than the computing logic itself while failing to dissipate localized hot spots. Daedalus uses a Direct Liquid Cooling (DLC) architecture operating on a closed water circuit.
Warm water is piped directly over copper cold plates sealed against the processor surfaces. Operating with higher inlet water temperatures dramatically reduces cooling energy overhead, allowing the facility to achieve an exceptionally low Power Usage Effectiveness (PUE) ratio even during peak summer ambient temperatures in Lavrio. Here is where most guides go wrong—they assume direct liquid cooling is strictly about power efficiency. In practice, stable thermal management prevents hardware thermal throttling, ensuring that multi-day distributed training jobs don't drop worker nodes mid-epoch.
The 60x Quantum Leap: Benchmarking ARIS Against Daedalus
To understand what 85.69 petaflops computing power in high performance computing actually delivers, compare it against Greece's legacy infrastructure. The nation's previous national system, ARIS (Advanced Research Information System), went live in 2015 delivering roughly 0.44 petaflops.
Daedalus increases national computational capacity by well over 100 times compared to ARIS's original spec, and vastly exceeds initial upgrade estimates. When the project was scoped under the EuroHPC Joint Undertaking framework in 2022, target figures sat around 30 to 60 petaflops. The final architecture exceeded those estimates through accelerated node hardware deployment and tuned interconnect topology.
| Specification / Metric | Legacy ARIS (2015) | Daedalus (2026 Launch) | European Peak Tier (e.g., JUPITER) |
|---|---|---|---|
| Sustained Performance (Rmax) | ~0.44 Petaflops | 85.69 Petaflops | > 1,000 Petaflops (Exascale) |
| Global TOP500 Rank | Unranked (>500) | Ranked 31st | Top 5 |
| Primary Accelerator | NVIDIA Tesla K40 / Phi | NVIDIA GH200 Grace Hopper | NVIDIA GH200 / Custom ASIC |
| Interconnect Topology | Infiniband FDR | High-Speed InfiniBand NDR | Custom Dragonfly / Slingshot |
| Cooling Mechanism | Forced Air / Standard Chillers | Direct Liquid Cooling (DLC) | Warm Water Direct Cooling |
| Primary Management | GRNET | GRNET / EuroHPC Consortium | EuroHPC / Jülich Supercomputing |
This performance leap changes the time-to-solution metric. A high-resolution coastal hydrodynamics model that previously took three weeks to run on ARIS can now finish during a single morning coffee break.
That said, there's a real catch here—raw FLOPS don't automatically guarantee application speedups. If software libraries aren't explicitly recompiled to take advantage of CUDA unified memory and MPI (Message Passing Interface) over high-speed InfiniBand NDR interconnects, legacy scientific code will leave 80% of the machine's capacity idling.
Sovereign Compute Infrastructure vs. Public Cloud Offloading
Why did the Greek government and the European Union commit €58.9 million to build a physical data center in Lavrio instead of simply renting compute instances on commercial clouds like AWS, Azure, or Google Cloud?
Commercial cloud environments excel at horizontal elasticity for web applications, but they carry severe structural trade-offs for scientific supercomputing:
- Data Egress and API Lock-in: Moving petabytes of raw earth-observation or genomic data out of hyperscaler environments yields prohibitive egress billing cycles.
- Network Latency across Distributed Workers: Massive parallel workloads require ultra-low latency inter-node communication. Hypervisor virtual layers in public clouds add microsecond-level delays that accumulate across tens of thousands of parallel core synchronized loops.
- Data Sovereignty and Compliance: National health databases, critical infrastructure telemetry, and sensitive civil protection archives carry strict regulatory constraints that prohibit storage on non-EU enterprise infrastructure.
By establishing a localized cluster under the GRNET national supercomputing infrastructure umbrella, the state secures guaranteed compute time at deterministic cost. The total expense covers physical site construction by MainSys, hardware delivery by HPE, and operational oversight by local research universities.
This hybrid model provides academic and public institutions with zero-cost access allocations, while keeping sensitive datasets strictly within national boundaries under European General Data Protection Regulation (GDPR) oversight. This next part matters more than it looks—it ensures early-stage scientific teams aren't priced out of foundational compute by fluctuating cloud credit costs.
Predictive Physics and Healthcare: Deploying Compute to Save Lives
Supercomputers earn their operational budgets when applied to complex physical systems. How does a supercomputer impact civil protection during natural disasters? Consider the operational reality of modeling summer wildfire dynamics across Mediterranean terrain.
Traditional wildfire models run simplified wind vector calculations over low-resolution elevation maps. They tell emergency teams where a fire might spread over the next four hours with moderate accuracy. Daedalus allows the Civil Protection Ministry to run dozens of high-resolution ensemble simulations simultaneously.
[Real-time Weather Data] + [Satellite Thermal Scanning] + [GIS Topography]
|routes into
[Daedalus HPC Compute Cluster]
|
+--------------------------+--------------------------+
| |
[Model A: High Wind Vector] [Model B: Humidity Shift Scenario]
| |
+--------------------------+--------------------------+
|generates
[Real-time Risk Heatmaps for First Responders within 10 Mins]
By processing micro-climate wind shifts, real-time drone telemetry, fuel load moisture content, and high-definition topography concurrently, the system updates risk maps every ten minutes. If wind directions shift over an island like Rhodes or Evia, disaster commanders receive predictive evacuation corridors before flames cross structural lines.
In healthcare, the platform acts as a diagnostic accelerator. The consortium—which includes IDIKA (Greece's e-Government Center for Social Security) along with institutions like the NCSR Demokritos and the Foundation for Research and Technology – Hellas (FORTH)—is structuring pipelines to run deep learning algorithms directly on regional medical imaging archives.
When a patient in a remote Aegean island health center receives a CT or MRI scan, local staff often lack immediate on-site sub-specialist radiologists. By processing those scans against deep learning algorithms trained on national medical image datasets, the system yields secondary diagnostic evaluations within minutes. This capability drastically reduces time-to-treatment for acute strokes and late-stage oncology flags.
The PHAROS AI Factory: Preserving Linguistic Equity in AI
Most modern Large Language Models (LLMs) are trained predominantly on English-language web crawls. As a result, low-resource and complex morphosyntactic languages like Greek suffer from tokenization inefficiencies, translation hallucination, and cultural context degradation.
To solve this, Greece is deploying the PHAROS AI factory Greece initiative directly on top of the Daedalus platform. PHAROS acts as an operational middleware layer: Daedalus provides the raw silicon, while PHAROS manages data curation pipelines, training frameworks, and model serving interfaces.
Training an LLM from scratch or fine-tuning massive open-weights models requires sustained matrix multiplication capabilities. Using the hardware's 85.69 petaflops capacity, Greek research teams can build domain-specific language models natively trained on legal text, administrative decrees, medical records, and historical literature.
This setup ensures three critical outcomes:
- Tokenization Efficiency: Native Greek tokenizers reduce the number of tokens required per sentence by up to 60% compared to standard GPT-4 tokenizers, directly cutting inference costs and latency for domestic software apps.
- Administrative Automation: Government agencies can deploy local digital assistants capable of processing complex administrative laws without sending citizen data across international boundaries.
- Cultural Preservation: Academic institutions can digitize, index, and analyze thousands of years of historical texts and dialect archives using native NLP models.
By controlling both the raw hardware infrastructure in Lavrio and the software framework via PHAROS, Greece avoids digital dependence on foreign tech monopolies for its core public sector AI software.
Frequently Asked Questions
How fast is the Daedalus supercomputer in Greece?
The Daedalus supercomputer Greece system achieves a measured benchmark performance of 85.69 petaflops (85.69 quadrillion floating-point operations per second). This computational output placed it 31st globally on the June 2026 TOP500 supercomputer rankings, making it over 100 times more powerful than Greece's prior ARIS cluster.
Who can access the Daedalus supercomputer Greece facility?
Access is provided free of charge to researchers, public universities, research centers, public health bodies, and startups. Resource allocations are managed by GRNET under peer-reviewed proposal calls, with dedicated compute time reserved for EuroHPC JU partner institutions across Europe.
How does a supercomputer impact civil protection during natural disasters?
A supercomputer processes high-resolution weather data, satellite imagery, and terrain maps simultaneously to run dozens of predictive scenario models. During wildfires or severe flooding, it projects hazard paths hours in advance, giving authorities actionable data to coordinate evacuations and deploy emergency teams.
Why was direct liquid cooling chosen for the Lavrio data center?
Direct liquid cooling removes thermal heat directly from the HPE processor modules using water loops, which is far more efficient than circulating chilled air. This architecture cuts total facility energy usage, allows denser server packing, and prevents thermal throttling during long artificial intelligence training cycles.
Building the Next Horizon of European HPC
Establishing sovereign compute capacity is no longer an abstract policy goal—it is a mandatory requirement for modern scientific research, environmental safety, and digital independence. Deploying the Daedalus supercomputer Greece facility equips researchers with the exact hardware tier needed to solve high-dimensional physical models and build native language AI systems locally. Assess your research group's computational requirements this quarter and submit your project proposal through the GRNET Allocation Portal. To explore how regional compute nodes are transforming public policy across the Mediterranean, read our technical overview of Sovereign Infrastructure in Europe.