GPU Fabric & Capacity
Phase-1: 8 servers × 8 NVIDIA B300 = 64 GPUs under one Kubernetes control plane
GPUs online
64 / 64
B300 · 180GB HBM3e
NVLink bandwidth
892 GB/s
92%
InfiniBand
348 Gb/s
87%
Memory bandwidth
2.1 TB/s
89%
Power draw
6.8 MW
▲ 5.1%
Avg temperature
61°C
within envelope
Cluster utilisation trend
GPU, memory, network and power
GPU fabric topology
NVLink mesh inside one node
GPU 0
84%
GPU 1
91%
GPU 2
98%
GPU 3
90%
GPU 4
97%
GPU 5
89%
GPU 6
96%
GPU 7
88%
NVLink 900 GB/s · InfiniBand 400 Gb/s · PCIe 128 GB/s · RoCEv2 / RDMA east-west fabric fully separated from management and IPMI networks.
Clusters
Capacity published to the Models page comes from these pools
| Cluster | GPUs | Utilisation | Memory | Temp | Active models | Throughput | Health |
|---|---|---|---|---|---|---|---|
| Cluster-A | 8 × B300 | 89% | 62°C | 24 | 612K t/s | Healthy | |
| Cluster-B | 8 × B300 | 82% | 58°C | 22 | 548K t/s | Healthy | |
| Cluster-C | 8 × B300 | 87% | 64°C | 26 | 587K t/s | Healthy | |
| Cluster-D | 8 × B300 | 84% | 60°C | 20 | 521K t/s | Healthy |
Bare metal pool
4 servers × 8 × B300 (32 GPUs) — Metal³ managed, customer dedicated infrastructure
Virtualisation pool
1 server × 8 × B300 (8 GPUs) — KubeVirt, SR-IOV networks, MIG / vGPU passthrough
KaaS / AI platform pool
3 servers × 8 × B300 (24 GPUs) — shared & dedicated inference, AI ISV workloads