GPU Cloud
Bare-Metal NVIDIA Clusters
Full NVLink domains on SXM nodes, dedicated interconnects, and isolated tenant networks. Built for serious AI workloads.
Bare-metal NVIDIA clusters with no noisy neighbours. SXM nodes with a full NVLink domain, giving distributed training jobs the GPU-to-GPU bandwidth the hardware was designed to deliver.
Full NVLink Domains
HGX B300 and H100 SXM nodes provide a complete NVLink domain, with no shared or partial domains. L40S nodes are PCIe-attached and are intended for inference, rendering and single-GPU workloads.
High-Performance Interconnects
HGX B300 clusters use RoCEv2 over 800G Ethernet with ConnectX-8. H100 SXM clusters are configured with InfiniBand or RoCEv2 depending on cluster generation and customer requirement. Final fabric configuration is confirmed at quote.
Isolated Tenant Networks
Per-tenant VLAN isolation and separate routing tables.
Kubernetes & Slurm
We support both Kubernetes and Slurm job scheduling.