Portfolio · Solution · Ignition line
You bring the science. Ignition hands back a working scientific capability, on a deadline: ten steps running from a first cloud experiment to your own AI factory. Each step is a result with a date, not a shipment of parts.
The ordering principle
Where the machine will live decides almost everything that follows: cooling, power, expansion, operation. That is why Ignition is organised into three environments: cloud, desk and field, datacenter. You recognise yours before looking at any product.
The ladder
One continuous ladder. Enter wherever your mission sits today; every rung above is a step, never a restart.
The first experiment, with no hardware.
Allocation on NVIDIA-certified cloud with GPU-hours included and scientific frameworks ready on day one. It ends with a documented step recommendation, the output isn't only scientific, it's the consumption profile that qualifies the next decision.
96 GB per GPU; up to 768 GB across eight concurrent GPUs.
Where the data cannot travel to the compute.
Field and edge inference module with 128 GB of unified memory. Operates outside a technical room: vehicle, vessel, greenhouse, factory floor. It doesn't compete with the other steps: it's often sold alongside one of them, in an end-to-end architecture.
Unified memory: a single coherent space.
The end of the cluster queue.
Individual development station with 128 GB of coherent memory. Two units can be linked over InfiniBand at 200G per port, totalling 256 GB aggregate: memory adds up, but each unit keeps its own 128 GB coherent domain. Usually the first hardware a group buys.
The ports are 200G InfiniBand. Memory across the two units adds up but does NOT form a coherent domain: a model that doesn't fit in 128 GB doesn't fit just because there are two machines.
Real multi-GPU, without a technical room.
Desk-side tower with up to four RTX PRO 6000 Blackwell Max-Q, 96 GB each. This is the bridge step: it solves the most common and most frustrating case in academic research, funding for compute, a building without infrastructure. Each card holds 96 GB coherent; aggregate reaches 384 GB.
Each GPU keeps its own coherent domain. Card-to-card communication is over PCIe; the total is aggregate, not coherent.
The capacity ceiling outside a datacenter.
Liquid-cooled desk-side station with 748 GB of coherent memory, the largest single domain available outside a technical room. It's the last step before the gate: when the client needs more, the conversation stops being about equipment and becomes about the facility.
The first rack step.
2U server based on the NVIDIA MGX reference architecture, with up to four RTX PRO 6000 Server Edition at 96 GB. It goes in alone and grows by adding nodes in the same rack, with no redesign. It's the line's commercial reference, the step against which the others are explained.
Each GPU keeps its own coherent domain. Card-to-card communication is over PCIe; the total is aggregate, not coherent.
Double precision in 2U.
Same 2U chassis, with two to four H200 NVL in 282 GB NVLink coherent domains. It isn't the expensive Beam: it's a product for a different problem. If the work requires FP64, the conversation is here or in Aperture, and the choice between them is about mesh size, not budget.
Coherent domains over NVLink. The aggregate is the sum of the domains.
Open PCIe. The next generation fits without replacing the server.
Open-PCIe 5U chassis, from two to eight accelerators. In the reference configuration, H200 NVL across up to two 564 GB coherent domains, the largest double-precision aggregate in the line. Every integrated-baseboard server freezes the GPU choice on purchase day; Aperture doesn't. For public-funded buyers, that's the difference between a three-year asset and a seven-year one.
Coherent domains over NVLink. The aggregate is the sum of the domains.
Native FP64 in the H200 reference configuration. With another accelerator, double precision follows the chosen GPU.
2.3 TB coherent in a single node.
Factory-integrated baseboard with eight B300 on a full NVLink fabric, the largest coherent domain in the line. It's the first step where the internal interconnect, not the GPU, is the product. Here the conversation stops being about equipment and becomes about a project.
Full NVLink fabric: all GPUs in a single coherent domain.
Here the product is the rack.
Eight nodes, eight B300 per node, 64 GPUs per rack, liquid cooling and 800G InfiniBand. Each node holds 2.3 TB of coherent memory; the rack aggregates 18.4 TB. ARETE isn't selling equipment: it delivers an environment, with network, cooling and commissioning as inseparable parts of the scope.
800G InfiniBand delivers very low-latency remote access, but NOT hardware cache coherence. The coherent domain ends at the node: 2.3 TB. The rack's 18.4 TB is aggregate.
Horizon
The destination, not a quote.
An engineering programme integrating every Hexa domain: power, cooling, network, compute, administration and assisted operation. Not an off-the-shelf product. It enters the conversation as a destination, always with two scenarios: as the client asked, and as ARETE recommends.
A complete facility, under project.
Expandability Index
The question no competitor answers in writing: how far does this take me before I have to replace everything.
| Level | Name | How it grows |
|---|---|---|
| 0 | Sealed | Fixed capacity at purchase. Growing means changing steps. |
| 1 | Internal | Grows inside the chassis, without replacing the system. |
| 2 | Pair | Aggregation limited to two units. |
| 3 | Horizontal in rack | Grows by adding nodes in the same room. |
| 4 | Fabric | Grows with no architectural ceiling. The building becomes the limit. |
| C | Elastic | Grows by allocation, not by hardware. |
The 3 envelopes
Step 01
Day-1 Frameworks, ALETHEIA, Continuity Path and one Pharos Session.
No physical asset: there is no CARE base and nothing to commission.
Steps 02–05
The above, with monthly Pharos. Praxis 0–1 is quoted separately, and whoever contracts it also receives the Acceptance Protocol and the STACK certificate.
Praxis 0–1 · Quoted separately
Steps 06–10
The above, with monthly Pharos. Praxis 0–1 is quoted separately, on-site, and whoever contracts it also receives the Acceptance Protocol and the STACK certificate.
Praxis 0–1 · Quoted separately · on-site
Deployment and Commissioning (Praxis 0–1) are quoted separately. Whoever contracts them also receives the Acceptance Protocol and the stack coherence certificate.
Data derived from catalogue ART-CAT-000 v2.3. Single public price: Cloud Start.
