AMD Instinct MI300X OAM Accelerator Tray (Dell PowerEdge XE9680)

The AMD Instinct MI300X OAM Accelerator Tray for the Dell PowerEdge XE9680 is built on AMD's CDNA 3 architecture, combining compute dies and HBM3 memory stacks into a unified Accelerated Processing Unit (APU) design. With 192 GB of HBM3 memory and a 5.3 TB/s aggregate memory bandwidth, the MI300X addresses the fundamental memory-capacity bottleneck that constrains large-scale AI model inference and training workloads. The OAM (OCP Accelerator Module) form factor enables dense multi-accelerator configurations within the XE9680 chassis, supporting up to eight MI300X modules per server node interconnected via AMD Infinity Fabric for high-bandwidth peer-to-peer communication.

At the compute level, the MI300X delivers 1,307 TFLOPS of FP16 performance and 2,615 TFLOPS of FP8 performance, making it well-suited for the numerical precision requirements of generative AI inference, large language model (LLM) training, and scientific simulation. The accelerator supports the ROCm open software platform, providing compatibility with widely adopted AI frameworks including PyTorch and TensorFlow, as well as HPC programming models such as HIP, OpenCL, and OpenMP offloading. ROCm's hardware abstraction layer enables enterprises to migrate and optimize existing CUDA-adjacent workloads without proprietary lock-in.

Deployed within the Dell PowerEdge XE9680 platform (manufacturer part number 210-BFZV), the MI300X OAM Accelerator Tray is an enterprise-grade component designed for datacenter rack integration with managed power, thermal, and serviceability characteristics appropriate for 24/7 mission-critical environments. The combination of the XE9680's system-level power delivery, NVMe storage backplane, and dual-socket CPU infrastructure with the MI300X's memory capacity makes this configuration particularly effective for serving very large foundation models—including those with hundreds of billions of parameters—entirely within a single server node's GPU memory pool.

Ideal for

• Large language model (LLM) inference serving, enabling multi-hundred-billion-parameter models to reside fully in-memory across the 8-accelerator memory pool of a single XE9680 node
• Generative AI training runs for transformer-based models where aggregate HBM3 capacity across accelerators eliminates the need for model parallelism across multiple nodes
• High-performance computing (HPC) workloads such as molecular dynamics simulation, computational fluid dynamics, and climate modeling that benefit from the CDNA 3 architecture's matrix and vector compute units
• Enterprise AI infrastructure consolidation, replacing multi-node CPU-based inference clusters with a smaller number of high-density GPU nodes to reduce rack space, power consumption, and network fabric complexity
• Mixed-precision scientific computing and AI-for-science applications that require both FP64 and lower-precision matrix operations within the same accelerator pipeline
• Retrieval-Augmented Generation (RAG) and embedding pipelines for enterprise search and knowledge management, where large embedding models and vector operations benefit from high memory bandwidth and capacity

Technical specifications

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