European accelerated-computing systems
Billi Dynamics designs GRIFO systems to keep more accelerators, data movement and control inside one coordinated local architecture.
GRIFO 32 · 32 GPUs per module · scale up to 128 — GRIFOTTO · 8 GPUs · available now
INSIDE GRIFO 32
Scale up within a single root complex.
The core remains the GRIFO 32 module, with 32 GPUs. By connecting up to four modules through the PCIe fabric, the architecture scales to 128 GPUs within a single root complex: multiple physical modules, one coordinated PCIe domain.
GRIFO, explained simply
GRIFO is a proprietary ecosystem of technologies designed by Billi Dynamics. Fabric, control, power and cooling keep accelerators working together, while qualified accelerators and I/O peripherals remain open to vendor choice.
Product family
GRIFOTTO: a complete 8-GPU system, configurable and orderable now. GRIFO 32: the 32-GPU core module of an architecture designed to connect up to four modules through PCIe fabric.
GRIFO 32 · FLAGSHIP SCALE-UP ARCHITECTURE
Single-root scale-up supernode
A 32-accelerator core module. The GRIFO PCIe fabric connects up to four modules to scale to 128 GPUs within a single root complex.
GRIFOTTO · AVAILABLE NOW
Compact 8-accelerator reference system
A transportable 8-GPU AI/HPC supernode built around Billi Dynamics system technologies, with dual-socket compute, high-speed I/O, management and full-system liquid cooling.
Choose the right system
Choose by workload, required scale and installation conditions. GRIFOTTO brings eight accelerators into one complete machine; GRIFO 32 starts with a 32-accelerator module and extends the local domain through its PCIe fabric.
Complete machine
GRIFOTTOA self-contained 8-accelerator system that can be configured and ordered now.Modular scale-up ecosystem
GRIFO 3232 GPUs per core module. Up to four modules linked by PCIe fabric within one root complex.| Layer | GRIFOTTO | GRIFO 32 |
|---|---|---|
| 01System identity | Complete systemA self-contained machine ready to host the workload. | Modular supernode ecosystemFabric, power, thermals and control are designed together; qualified accelerators and I/O remain selectable. |
| 02Local accelerator domain | 8 Gen5 x16 GPU positionsA dual-socket, dual-root NUMA domain. | 32 accelerators per module · up to 128 GPUsUp to four GRIFO 32 modules connected through PCIe fabric within one root complex. |
| 03Fabric | Dedicated Gen5 x16 connectionsInternal PCIe paths defined and validated at machine level. | Single-root switched PCIe Gen5 fabricConnects accelerators within each module and extends the shared PCIe domain between modules. |
| 04Scale-out network | 1 × 400G-ready NIC/HCA positionConnects storage or external infrastructure according to the deployment. | Scale up locally, then scale outAdditional high-speed NICs/HCAs connect already-capable supernodes, storage and the data centre. |
| 05Cooling architecture | Full-system liquid coolingCPU, accelerators and critical components are managed inside the same machine. | Three-layer thermal architectureCold plates, distribution/CDU and facility heat rejection are sized with the domain. |
| 06Power architecture | Integrated machine power distributionConfigured around the number and envelope of the installed accelerators. | Domain-level power deliveryDistribution, protection, cabling and thermal load are coordinated across the ecosystem. |
| 07System control | BMC, firmware and topology validationComplete operational control of the individual system. | Coordinated domain controlEnumeration, switching, firmware, telemetry and validation span every subsystem. |
| 08Deployment model | A standalone machine or cluster nodeFor labs, AI/HPC, simulation and compact deployments. | A data-centre building blockEnlarge the local machine first; then connect complete GRIFO domains at infrastructure scale. |
| 09Status | Configurable and orderable nowA documented reference system undergoing real validation. | Flagship architectureIndustrialisation and validation of the 32-accelerator domain. |
“Start with 32 GPUs. Connect up to four GRIFO 32 modules through PCIe fabric. Scale to 128 GPUs within one root complex.”
The definition
01The PCIe fabric connects accelerators and GRIFO 32 modules into a shared local domain. Pooling, I/O management and communication topology define how resources are reached and coordinated.
02Power delivery and full-system liquid cooling are designed alongside that data path. High-performance networking provides the scale-out connections to storage and other computing domains.
Select qualified accelerators and I/O peripherals for the workload, software stack and procurement requirements. The GRIFO system architecture provides continuity as those device choices evolve.
BILLI DYNAMICS CORE · QUALIFIED ACCELERATORS + I/OFROM NETWORKED SERVERS TO ONE COORDINATED DOMAIN
Every server has its own boundary. Data must repeatedly cross an external packet network.
Independent hosts · external network between nodes
Accelerators, fabric and control are engineered as parts of one larger local machine.
32-GPU module · up to 128 GPUs by linking four modules through the PCIe fabricIllustrative renderings: networked servers on the left, a single GRIFO 32 module on the right. The equal-accelerator comparison follows in the next section.
Accelerators stop behaving like isolated resources in separate servers and become parts of one coordinated computing domain.
Because compute, memory, fabric, control, power and cooling are designed as parts of one machine.
AI is increasingly limited by data distance, energy and heat — not only by the speed of the chips.
As models and datasets grow, moving data between isolated servers becomes more expensive. The architectural direction is therefore unavoidable: enlarge the useful local machine before scaling out across the data centre.
Architecture study · 128-accelerator reference scenario
The study gives the conventional cluster 15.6% more theoretical compute. Then it follows what the data must cross before that compute can become useful work.
A deliberately conservative starting point
Both reference systems contain 128 NVIDIA H200 accelerators, using the form-factor and power variants stated in the study. The comparison assumes an idealised, perfectly tuned 1:1 fat-tree for the cluster and makes no claim that peak FLOPS equal application performance.Independent hosts · external network between nodes
Four GRIFO 32 modules · PCIe fabric · one root complex
Illustrative architecture renderings. The comparison uses the configurations stated above; the number of units pictured is not a literal representation of the study topology.
Follow one exchange
MODELED COMMUNICATION PATH≥850 ns physical floor
≈505 ns switched local path
The reference model places GRIFO’s local switched path at about 505 ns, compared with a conventional network path whose physical components alone total at least 850 ns.
The local exchange no longer needs to be serialised through a NIC, converted to an optical link, switched as a network packet and reconstructed at another server.
A component advantage on paper can be overtaken by time spent moving and coordinating data. The complete architecture determines how much compute reaches the workload.
“Peak compute is the starting number. Boundaries decide how much reaches the workload.”
Figures are derived from the stated 128-accelerator analytical reference model. They are not measured GRIFO 32 application results. Actual performance depends on the exact accelerator, topology, software stack and workload.
Business case · 8,192-GPU iso-topology
This analytical case compares 256 GRIFO 32 modules with 1,024 conventional 8-GPU nodes. Both contain 8,192 GPUs; each node has 8 × 400G uplinks. The estimates below apply to this specific infrastructure and operating scenario.
H200 accelerators in both architectures
GRIFO 32 domains · 32 GPUs each
conventional DGX nodes · 8 GPUs each
uplinks per node · non-blocking fat-tree
lower at the midpoint of the published estimate ranges
€400–485M conventional · €305–390M GRIFO 32lower across energy, MRO, people and spares
€132M conventional · €56M GRIFO 32lower at the same 8,192-GPU scale
≈30% below the conventional cluster modelThe curve applies the published 8,192-GPU TCO result proportionally to the deployment checkpoints. It shows the economic effect accumulating as GRIFO replaces more conventional server boundaries; intermediate values are interpolation, not independent quotations.
The model separates the capital needed to build the infrastructure from the operating cost required to keep it running.
wall power at 8,192 GPUs
energy avoided over five years
energy cost avoided at €0.15/kWh
Start with the same cost categories: hosts, NICs, switches, optics, power and cooling. For one or a few machines, the advantage must be calculated for the actual configuration and utilisation. The 8,192-GPU case illustrates the effect at scale; it does not determine your project's savings.
“The value accumulates as more repeated server infrastructure is consolidated.”
Source: GRIFO 32 Cluster 8,192 GPU — Business Edition Rev. 1.1, April 2026. Figures are order-of-magnitude estimates for the stated NVIDIA H200 iso-topological reference case, 24/7 full load, €0.15/kWh and a five-year horizon. CAPEX benefit uses the midpoint of the published ranges; intermediate scale values are proportional interpolation. Actual costs depend on configuration, workload, procurement, facility, utilisation and energy price.
Software platforms
Choose the tools for your data, models and business processes.
Move computation to the data instead of moving data through layers of infrastructure. GRIFO Data Engine brings relational, vector and AI-assisted operations closer to accelerator memory, turning large working sets into faster, more efficient decisions.
Best forLarge analytical working sets, vector search and AI-fused enterprise data.
Business advantages
Turn confidential archives into usable institutional intelligence—inside your security perimeter. ConfidentialGPT helps teams find evidence, compare documents, reveal inconsistencies and prepare decisions without exporting sensitive knowledge to an external cloud.
Best forLegal, finance, public sector, research and organisations handling sensitive knowledge.
Business advantages
Move from reports that describe the past to scenarios that illuminate the next decision. PROFETA combines forecasting, simulation and operational context so organisations can test alternatives before their consequences become costs.
Best forEnergy, manufacturing, operations and planning teams.
Business advantages
Give operational systems eyes and ears. iVST converts video, images and complex sensor signals into correlated events that inspection, security and automation systems can understand and act upon.
Best forIndustrial inspection, multi-camera analytics, security and autonomous systems.
Business advantages
Why architecture matters
A peak specification describes what one component could do under ideal conditions. Real work finishes only when data reaches the accelerators, memory keeps them fed, the fabric coordinates them, software uses them efficiently, and power and cooling sustain the machine for the entire run.
What the accelerator could compute when every supporting condition is ideal.
A NUMBER ON A SPEC SHEETHow much correct work the complete system actually finishes per unit of time, energy and cost.
A MEASURABLE RESULTIf data, memory, fabric, software or cooling slows down, the accelerators wait. Adding more peak compute does not remove that waiting time; the complete path must be redesigned.
The machine must match the model, precision, batch size and required time-to-result.
Data must already be close to the compute instead of waiting beyond distant server boundaries.
Capacity keeps the problem resident; bandwidth feeds accelerators at the rate they consume.
Every exchange and synchronisation needs enough bandwidth, low enough latency and predictable behaviour.
Drivers, runtimes and scheduling must expose the hardware as one coordinated system.
The facility must sustain clocks and density for the entire run, not only for a short burst.
Documented evidence
Reference configurations and validation scope are stated separately from analytical studies.
8 Gen5 x16 GPU positions · dual Xeon reference compute · 2 TB reference memory · 400G-ready I/O
Up to 32 full-length PCIe accelerators · one single-root Gen5 switched fabric · liquid cooling by design
Power · firmware · BMC · GPU/PCIe/NUMA topology · network and GPUDirect checks
BENCHMARK POLICYApplication benchmark results will be published only with the exact configuration, workload, software versions and measurement conditions.
Solutions
We define the workload, data path, facility boundary and success criteria before configuring the machine.
Knowledge, models and workflows under direct organisational control.
CONFIDENTIALGPT + GRIFOMore accelerators in a local domain, with thermals and facility designed together.
GRIFO 32Simulation, accelerated analytics and AI in a workload-oriented system.
GRIFOTTO + GRIFORelational, vector and AI operations closer to accelerator memory.
GRIFO DATA ENGINE
Founder & System Architect
For more than twenty years, between Europe and the United States, Emilio Billi has worked on HPC, PCIe scale-up, shared memory, accelerated systems and hardware-software co-design. His method starts from the path of data and turns physical limits into design decisions.
“The system is the product. Measurement is the proof.”
Company
Billi Dynamics S.r.l. turns system architecture into industrial products for AI, scientific computing, analytics and mission-critical deployments.
ENGINEERING PROOFReal system assembly and validation in the Billi Dynamics laboratory.
Leadership and board
Founder & CTO (acting CEO)
Deep-tech entrepreneur and system architect with more than 20 years across HPC, AI infrastructure and hardware-software co-design.
BizOps
Connects business operations, strategy and execution to support controlled scale-up.
Board Member
Industrial leader with experience in innovation, digitalisation, operational excellence and growth.
CSO
Commercial and strategy leader for positioning, alliances and revenue-focused growth.Start with your workload
Tell us about your application and the scale you need. The first technical conversation helps define a suitable configuration, installation requirements and the validation path for your workload.
Define your data, models, capacity and time-to-result requirements.
Review the system, accelerators, I/O, power and cooling requirements.
Agree which tests are needed and the installation, commissioning and support scope to include in the proposal.
MODEL · NUMBERS · MEASUREMENT · DECISION
