Several servers joined by a fast network
Three things to consider before comparing the specifications below.
Choose an architecture for simulation, AI, private cloud or storage, then size the node count.
Match the interconnect to how much data your jobs exchange between machines.
Size shared storage for datasets, checkpoints and results, alongside your recovery needs.
The four architectures, each at both published sizes, as specified and tested.
| System |  Fusion Grid Sim 200 |  Fusion Grid Sim 300 |  Fusion Grid AI 200 |  Fusion Grid AI 300 |  Fusion Grid Cloud 200 |  Fusion Grid Cloud 300 |  Fusion Grid Vault 200 |  Fusion Grid Vault 300 |
|---|---|---|---|---|---|---|---|---|
| Overview | Four-node simulation cluster | Sixteen-node simulation cluster | Four-node training cluster | Eight-node training cluster | Nine-node private cloud | Twenty-five node private cloud | Six-node storage cluster | Storage cluster: 12 storage nodes plus 3 metadata nodes |
| Architecture | Engineering simulation | Engineering simulation | AI and ML training | AI and ML training | Private cloud | Private cloud | Distributed storage | Distributed storage |
| Nodes | 4 | 16 | 4 | 8 | 9 | 25 | 6 | 12 |
| Cores | 256 | 1,536 | — | — | 384 | 1,536 | 384 | 1,152 |
| Accelerators | 4 × 32GB | 32 × 48GB | 4 × 48GB | 16 × 48GB | 2 × 48GB | 8 × 48GB | None | None |
| Interconnect | HDR 200Gb/s | NDR 400Gb/s | 2 × HDR 200Gb/s | 2 × NDR 400Gb/s | 2 × 25GbE | 2 × 100GbE | 2 × 25GbE | 2 × 100GbE |
| Storage | 150TB Lustre | 500TB Lustre | 100TB NVMe | 400TB NVMe | 120TB Ceph | 600TB Ceph | ~1.05PB | ~4.6PB |
| Cabinets | 1 | 2 | 1 | 1 | 1 | 2 | 1 | 2 |
| Action |
The architectures we have built and tested, each at the size it was designed for and at the smallest version that is still the same architecture. Quoted rather than priced, because a cluster is sized to the room it goes into. The nodes inside them are the servers we sell on their own.
MPI jobs such as ANSYS Fluent and MATLAB split one problem across many machines, and every step waits for the slowest link between them. So the network decides how far a case scales, long before core count does. Memory speed is the second limit, because most scientific codes run out of memory bandwidth before they run out of cores. Built on OpenHPC and Slurm, with the network sized so that scaling is limited by the software, not the hardware.
The same OpenHPC stack, scheduler and way of working as the 300, on four nodes and an HDR network at two-to-one.
Sixteen dual-socket Turin nodes, 1,536 cores on fully filled memory channels.
AI and machine learning need different things from a simulation cluster. Accelerator memory and speed matter more than processor cores. The processors prepare data, and the accelerators do the work. Video memory usually decides everything, because a model that does not fit does not run slowly. It does not run at all. The network is sized so parameters and gradients move without holding training back, and the data pipeline is sized to keep the cards busy, since a slow loader leaves your most expensive hardware idle.
Four nodes with one L40S each, pooling 192GB, on HDR InfiniBand, or 100GbE RoCEv2 where a site already runs Ethernet.
Sixteen L40S across 8 nodes, pooling 768GB of accelerator memory.
A private cloud gives you cloud computing on hardware you own and run. MAAS manages the physical machines (PXE discovery, IPMI power control and automated operating system deployment), and OpenStack adds virtual machines, software-defined networking and storage on top. Performance depends as much on the control plane staying available, and on data moving efficiently between components, as on processor speed. That is why the controllers are specified separately from the compute. Choose it for data residency, security or regulatory needs, and for workloads where owning the hardware costs less than renting it.
Still 3 controllers, because two cannot form a quorum and the control plane is the reason to own this instead of renting.
Three dedicated controllers, 12 compute nodes, 4 accelerated nodes and 6 Ceph OSD nodes.
A distributed storage cluster spreads data across nodes, so capacity, speed and redundancy all grow as you add hardware. Ceph, Lustre and MinIO coordinate where data lives, how it is accessed and how it recovers, and keep serving while individual drives or whole nodes are missing. Processing power matters most during erasure coding and recovery, and metadata performance is what usually limits a real workload, so metadata is kept apart from bulk capacity here.
Six 2U nodes with 12 bays each, a petabyte usable at four-plus-two, with metadata on the same nodes instead of separate hardware.
Twelve 4U nodes with 36 bays each, 6.3 petabytes raw and 4.6 usable.
Tell us what has stopped fitting in one machine, such as the solver, the model or a queue that no longer clears overnight. We will size one of these to it and come back with the same verification report every other machine on this site carries.
A tower that sits beside your desk on a normal power socket and is quiet enough for an office. One person uses it and looks after it.
8 machinesA two-processor server in a rack, with backup power supplies and 24 filled memory channels. Built for work that has to keep running while nobody is watching.
8 machines
