HPE ProLiant XL675d Gen12

The HPE ProLiant XL675d Gen12 is a next-generation tray-form server designed from the ground up for high-density GPU-accelerated AI workloads at hyperscale and enterprise scale. Built on the AMD EPYC Turin (9005 Series) processor architecture, the XL675d Gen12 delivers substantial per-socket compute headroom with support for up to 192 cores per socket, ensuring that CPU-side data preprocessing, orchestration, and model management keep pace with GPU throughput demands. Each tray accommodates up to 8 high-performance GPUs—compatible with NVIDIA H100 and next-generation accelerators—connected via PCIe Gen 5 fabric with NVLink bridge support, enabling tight GPU-to-GPU communication essential for large-scale transformer model training and inference pipelines.

The system is engineered for deployment inside HPE's Apollo and Cray SuperDome liquid-cooling chassis ecosystems, leveraging direct liquid cooling (DLC) to sustain full GPU and CPU thermal design power under continuous training workloads. High-bandwidth DDR5 memory channels paired with CXL 2.0 memory expansion capabilities allow memory capacity to scale well beyond traditional DIMM ceilings, addressing the working-set demands of foundation model training. The XL675d Gen12 integrates HPE's iLO 7 baseboard management controller for deep telemetry, predictive health analytics, and GPU power capping, enabling fleet-level orchestration across thousands of nodes in an AI factory topology.

Noting its 2025-Q4 launch, the XL675d Gen12 reflects HPE's architectural strategy for the AI infrastructure era: disaggregated, liquid-cooled, and software-defined at every layer. Front-end storage connectivity via NVMe-oF and rear-fabric high-speed interconnects (400GbE / InfiniBand NDR) ensure that storage and networking bandwidth do not bottleneck GPU utilization. The platform is validated with HPE's Machine Learning Development Environment (MLDE) and integrates natively with HPE Cray EX supercomputing software stacks, making it equally suited to national-lab-scale AI clusters and enterprise AI factories.

Ideal for

• Large-scale generative AI and large language model (LLM) pre-training requiring sustained all-to-all GPU communication across hundreds of tray nodes
• Computer vision and multimodal foundation model training in hyperscale data centers where rack density and power efficiency are critical operational constraints
• Enterprise AI factory deployments consolidating training, fine-tuning, and inference serving within a thermally optimized liquid-cooled chassis environment
• High-performance scientific simulation and HPC workloads co-located with AI pipelines, leveraging AMD EPYC Turin's high core count for CPU-intensive preprocessing stages
• Reinforcement learning from human feedback (RLHF) and distributed fine-tuning pipelines requiring rapid GPU-to-GPU gradient synchronization via NVLink and high-speed fabric
• National research and defense AI workloads demanding sovereign, on-premises training infrastructure with enterprise-grade management via HPE iLO 7 and Cray system software

Technical specifications

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