HPE Cray XD675 AI Accelerator Server

The HPE Cray XD675 is a purpose-built 4U AI accelerator server designed within the Open Compute Project (OCP) Data Center Ready form factor, engineered to support the most demanding large language model (LLM) training, fine-tuning, and high-performance computing workloads. At its core, the XD675 accommodates up to eight NVIDIA H200 SXM5 Tensor Core GPUs interconnected via fifth-generation NVLink, delivering a combined GPU memory capacity of up to 1.12 TB of HBM3e at 3.35 TB/s of aggregate memory bandwidth. This architecture eliminates inter-GPU communication bottlenecks, enabling near-linear scaling across transformer model training tasks that demand rapid gradient synchronization and massive parameter counts.

The XD675 is designed for datacenter-scale deployment within the HPE Cray supercomputer ecosystem, integrating tightly with HPE's high-speed Slingshot 11 interconnect fabric to enable multi-node scaling across thousands of GPU endpoints. Dual fourth-generation AMD EPYC processors provide the CPU compute backbone, supplying ample PCIe Gen 5 lanes, high memory bandwidth, and the I/O throughput necessary to keep eight SXM5 GPUs fully saturated with training data. The server supports up to 6 TB of DDR5 system memory, and its NVMe storage subsystem ensures rapid checkpoint saving and dataset streaming without creating I/O-bound bottlenecks during training runs.

Direct liquid cooling (DLC) is a first-class design element in the XD675, with chassis-integrated cold plates covering both GPUs and CPUs to sustain peak thermal dissipation at datacenter scale without reliance on airflow-intensive hot-aisle containment alone. This approach enables higher rack densities and aligns with modern sustainable datacenter power and cooling strategies. The platform is managed through HPE iLO 6 with integrated Redfish API support, enabling programmatic lifecycle management, telemetry streaming, and integration with HPE GreenLake cloud management services for unified fleet observability across on-premises AI infrastructure.

Ideal for

• Large language model pre-training and fine-tuning at scale, leveraging NVLink-connected H200 SXM5 GPU clusters spanning hundreds of nodes via HPE Slingshot fabric
• Generative AI inference serving for latency-sensitive enterprise applications requiring massive GPU memory capacity to load multi-hundred-billion-parameter models in full precision
• Multimodal AI model development combining vision transformers and language models that benefit from the 1.12 TB aggregate HBM3e memory pool across eight GPUs
• Scientific simulation and computational fluid dynamics workloads in national laboratories and research institutions requiring sustained FP64 double-precision throughput
• Drug discovery and molecular dynamics simulations where GPU-accelerated quantum chemistry and protein folding models demand extreme memory bandwidth and capacity
• Sovereign AI and enterprise AI cloud deployments where organizations require on-premises GPU infrastructure with full datacenter integration and programmatic management via Redfish and HPE GreenLake

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.