Dell PowerEdge XE9680 8x MI300X OAM Server – CTO GPU Node

The Dell PowerEdge XE9680 configured with eight AMD Instinct MI300X OAM accelerators represents one of the most memory-dense and compute-capable GPU server platforms available for enterprise AI and HPC deployments. The XE9680 chassis is purpose-built to house OAM (Open Accelerator Module) form-factor GPUs, and the MI300X variant brings AMD's CDNA 3 architecture to bear with a unified CPU+GPU memory architecture—each MI300X integrates 192 GB of HBM3 memory, yielding a total of 1.5 TB of high-bandwidth accelerator memory across all eight GPUs in a single 5U node. This extraordinarily large unified memory pool enables workloads that would otherwise require model sharding across multiple nodes to run within a single server, dramatically reducing inter-node communication overhead.

The XE9680 platform supports PCIe Gen 5 interconnects and is engineered around high-throughput AMD Infinity Fabric links between accelerators, enabling rapid all-to-all GPU communication suited to large-scale model parallelism. The system accommodates high-core-count AMD EPYC (Genoa-class) processors, high-capacity DDR5 system memory, and NVMe storage expansion, making it a fully self-contained AI training and inference node. Liquid cooling readiness and a dense 5U form factor address the thermal and spatial challenges inherent in deploying this class of accelerator at data center scale.

As a Configure-To-Order (CTO) offering, the XE9680 MI300X GPU Node allows enterprise procurement teams to tailor CPU, DRAM, storage, networking, and firmware configurations to specific workload and infrastructure requirements. This platform is well suited to large language model training, generative AI inference, high-performance scientific computing, and GPU-accelerated data analytics pipelines in UAE, GCC, EMEA, and APAC enterprise and hyperscale data center environments.

Ideal for

• Large language model (LLM) pre-training and fine-tuning requiring multi-hundred-billion parameter models to reside entirely within on-accelerator HBM3 memory without multi-node sharding
• Generative AI inference serving for production deployments of foundation models where low-latency, high-throughput responses are required at enterprise scale
• High-performance scientific computing and simulation workloads in energy, life sciences, and engineering that require both dense compute throughput and large fast-access memory capacity
• GPU-accelerated data analytics and machine learning pipelines processing large in-memory datasets for financial services, telecommunications, and government sectors
• Multi-tenant AI-as-a-Service infrastructure within private or sovereign cloud data centers serving GCC and APAC regulated industries
• Retrieval-augmented generation (RAG) and vector database acceleration where large embedding models and indices must be held in high-bandwidth accelerator memory for real-time query performance

Technical specifications

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