HPE ProLiant Compute XD685 with AMD Instinct MI325X

The HPE ProLiant Compute XD685 is a purpose-built, high-density GPU accelerator node designed around eight AMD Instinct MI325X accelerators, delivering an unprecedented combination of compute throughput and memory bandwidth for the most demanding AI and HPC workloads. Each MI325X accelerator integrates 256 GB of HBM3E memory per GPU, yielding a total of 2 TB of ultra-high-bandwidth memory across the full eight-GPU configuration. This extraordinary memory capacity directly addresses the growing model-size requirements of large language models (LLMs), multimodal foundation models, and transformer-based architectures that previously required complex multi-node sharding strategies.

Built on the HPE ProLiant XD family architecture, the XD685 leverages a high-bandwidth interconnect fabric between accelerators, enabling near-native GPU-to-GPU communication with minimal latency. The platform supports PCIe Gen 5 connectivity and is designed to integrate with high-speed InfiniBand and Ethernet networking fabrics, enabling scale-out clusters capable of training trillion-parameter models. The node features advanced liquid-cooling readiness and a dense 4U form factor optimized for modern AI data centers where power and space efficiency are critical operational priorities.

Launched in Q2 2025, the HPE ProLiant Compute XD685 positions organizations to accelerate generative AI development cycles by reducing per-iteration training time and enabling larger batch sizes. HPE integrates the system with its broader software ecosystem including HPE Machine Learning Development Environment (MLDE) and HPE Cray ClusterStor for storage convergence, providing an end-to-end validated infrastructure stack. The platform is certified for enterprise deployment in hyperscale AI factories, national research computing centers, and large-scale enterprise AI programs where sustained throughput and system reliability are non-negotiable.

Ideal for

• Large language model (LLM) pre-training and fine-tuning for enterprise generative AI applications requiring multi-terabyte memory capacity per node
• Multimodal foundation model development combining vision, language, and audio modalities that demand massive GPU memory to hold full model replicas during training
• High-throughput AI inference serving for simultaneously deployed large-scale models in production environments requiring low-latency, high-concurrency responses
• Scientific simulation and computational fluid dynamics workloads in national labs and research institutions leveraging AMD ROCm-optimized HPC libraries
• Retrieval-augmented generation (RAG) pipeline acceleration where large embedding models and vector operations benefit from HBM3E memory bandwidth
• Distributed reinforcement learning from human feedback (RLHF) training workflows requiring dense GPU memory to store critic and actor model states concurrently

Technical specifications

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