Skip to main content
KONST

PT-B300

Portugal
NVIDIA B300 SXM cluster for large pretraining workloads

A European facility with 128 nodes and 1,024 B300 SXM GPUs, 288 GB of memory per GPU, and 800 Gbps RoCE v2 for LLM pretraining and HPC.

LLM pretrainingInferenceHPC

1,024

Total GPUs

128

Nodes (8x B300 per node)

800Gbps

GPU Fabric (RoCE v2)

4.55 and up

USD/hr/GPU (3 or 5 year term)

Compute nodes

128 nodes with 8 NVIDIA B300 SXM GPUs per node.

GPU
1,024 units · 8x B300 SXM/node
GPU memory
288 GB HBM3e per GPU, 2,304 GB per node, 288 TB total
CPU
2x Intel Xeon 6767P per node (128C/256T)
Memory
2 TB DDR5 per node (DIMM configuration available on request)
Storage
OS: 2x boot SSD (details available on request); data: 8x 3.84 TB NVMe (30.72 TB per node, approximately 3.93 PB raw across the cluster)

Network and cooling

800 Gbps RoCE v2 GPU Fabric.

Network (GPU Fabric)
800 Gbps Ethernet RoCE v2, 8x ConnectX-8 per node (6.4 Tbps per node)
Network (External)
2x 400 GbE per node; 10 Gbps external internet
Cooling
Contact sales

Note

Prices are indicative hourly rates per GPU and vary by contract length, memory option, and payment structure. Some facility cities, data center names, Tier ratings, and contracting entities are available on request and will be disclosed during contracting. The formal quote sets the final terms.

Planning LLM pretraining at 1,024 GPU scale?

We can provide the full network topology, configuration details for 288 TB of GPU memory and 800 Gbps RoCE v2, and an estimated price range.

Book a consultation