HPE ProLiant Compute DL384 Gen12

The HPE ProLiant Compute DL384 Gen12 represents a new class of AI-optimized infrastructure, engineered around the NVIDIA GH200 Grace Hopper Superchip architecture. Each GH200 module tightly integrates a 72-core ARM Neoverse V2 Grace CPU with an NVIDIA Hopper H100 GPU via a high-bandwidth NVLink-C2C interconnect delivering 900 GB/s of bidirectional bandwidth, effectively eliminating the traditional PCIe bottleneck that constrains GPU memory-access performance in conventional server designs. The DL384 Gen12 houses multiple GH200 Superchips in a dense, thermally engineered 4U chassis with direct liquid cooling support, enabling sustained full-performance operation across all accelerators simultaneously without thermal throttling.

Designed explicitly for generative AI, large language model (LLM) training, and high-throughput inference workloads, the DL384 Gen12 leverages the massive unified HBM3e memory pool available per GH200 module—up to 96 GB HBM3e per GPU combined with up to 480 GB LPDDR5X CPU memory per Superchip—providing an unprecedented addressable memory space that enables in-memory operation of multi-hundred-billion parameter models without model sharding across nodes. NVSwitch fabric interconnects across multiple Superchips within the system deliver NVLink-based all-to-all GPU communication, dramatically accelerating collective operations such as AllReduce that are critical in distributed training runs.

HPE integrates the DL384 Gen12 with its ProLiant ecosystem management stack including iLO 7 with AI-assisted telemetry, HPE Integrated Smart Update Manager, and native support for HPE GreenLake cloud services for hybrid AI infrastructure management. The server is validated for deployment within HPE Cray supercomputing clusters as well as standalone rack configurations, supporting NVIDIA Base Command Manager, NVIDIA AI Enterprise software suite, and major MLOps platforms. Its front-to-rear airflow and direct liquid cooling (DLC) architecture make it suitable for high-density AI data centers targeting Power Usage Effectiveness (PUE) efficiency at scale.

Ideal for

• Training and fine-tuning large language models with 70B+ parameters entirely within unified GPU-CPU HBM3e memory without multi-node model parallelism overhead
• High-throughput generative AI inference serving for enterprise LLM APIs, reducing token latency through near-zero PCIe memory transfer bottlenecks
• Scientific simulation and high-performance computing workloads requiring tightly coupled CPU-GPU compute such as molecular dynamics, climate modeling, and computational fluid dynamics
• Multimodal AI model development including vision-language models and diffusion-based image and video generation requiring large GPU memory footprints
• Enterprise AI platform deployments running NVIDIA AI Enterprise on-premises with HPE GreenLake hybrid cloud management for secure, governed model development pipelines
• Retrieval-augmented generation (RAG) pipeline acceleration where large vector databases and LLM inference must co-reside in memory for sub-millisecond retrieval-to-generation latency

Technical specifications

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