HPE ProLiant XD685 Gen11 8x MI300X GPU Server

The HPE ProLiant XD685 Gen11 is a purpose-built, high-density GPU server engineered to accelerate large-scale AI training, inference, and HPC workloads at the datacenter level. The system integrates eight AMD Instinct MI300X Accelerated Processing Units, each incorporating AMD's unified memory architecture that combines GPU compute dies and HBM3 memory stacks into a single package, delivering exceptional memory bandwidth and capacity per accelerator. Designed around the HPE ProLiant Gen11 platform, the XD685 supports dual AMD EPYC 9004 series processors and a high-throughput interconnect fabric that ensures efficient data movement between host CPUs and the GPU array.

The MI300X accelerators within this server are equipped with 192 GB of HBM3 memory per GPU, yielding an aggregate on-accelerator memory pool of 1.5 TB across the eight-GPU configuration. This extraordinary memory capacity makes the XD685 Gen11 particularly well suited for serving very large language models (LLMs) entirely within accelerator memory, eliminating the latency penalties associated with memory offloading. The system leverages high-bandwidth GPU-to-GPU interconnects and PCIe Gen 5 host connectivity to sustain the throughput demanded by foundation model training and real-time generative AI inference pipelines.

From a datacenter operations perspective, the HPE ProLiant XD685 Gen11 is designed for rack-scale deployment with HPE's OpenBMC-based iLO 6 management controller, enabling out-of-band monitoring, firmware lifecycle management, and integration with HPE GreenLake cloud-based infrastructure management. The server's mechanical design targets high-density thermal management appropriate for the power envelopes associated with eight MI300X accelerators, supporting enterprise IT teams responsible for large AI infrastructure estates across cloud, colocation, and on-premises datacenter environments.

Ideal for

• Large language model (LLM) training and fine-tuning for enterprise generative AI initiatives requiring massive GPU memory capacity
• High-throughput AI inference serving of foundation models such as GPT-class and multimodal architectures that benefit from keeping full model weights resident in accelerator memory
• Scientific computing and HPC simulation workloads in sectors such as energy, life sciences, and engineering that demand sustained FP64 and FP32 compute alongside high memory bandwidth
• Retrieval-augmented generation (RAG) pipelines and vector database acceleration supporting enterprise AI applications at scale
• Multi-tenant AI platform deployments where datacenter operators require dense GPU consolidation to maximize accelerator utilization per rack unit
• Computer vision model training and video analytics inference at scale for smart infrastructure, manufacturing quality control, and security applications

Technical specifications

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