AMD Instinct MI300X 8-GPU OAM Server (HPE Cray XD670)

The AMD Instinct MI300X is a data center GPU accelerator built on AMD's CDNA 3 architecture, integrating a unified GPU compute die with high-bandwidth memory in a single Accelerated Processing Unit package. The MI300X features 192 GB of HBM3 memory with an aggregate bandwidth of 5.3 TB/s, making it one of the highest-memory-capacity GPU accelerators available for large-scale AI inference and training. In the HPE Cray XD670 server configuration (HPE part number R9G71A), eight MI300X OAM (OCP Accelerator Module) units are installed in a dense 2U chassis, providing a combined 1.536 TB of GPU memory per node and enabling the execution of very large language models and multi-modal AI workloads without model sharding across nodes.

The CDNA 3 architecture delivers significant improvements in matrix math throughput compared to prior generations, supporting FP64, FP32, FP16, BF16, FP8, and INT8 precisions natively. The MI300X OAM modules communicate via high-speed Infinity Fabric interconnects within the HPE Cray XD670 chassis, enabling coherent GPU-to-GPU data exchange at low latency. The HPE Cray XD670 platform is purpose-built for dense GPU deployments, featuring a high-efficiency liquid cooling infrastructure and PCIe Gen 5 host connectivity to support the bandwidth demands of the MI300X accelerators.

The AMD Instinct MI300X 8-GPU OAM Server targets enterprise data centers, national laboratories, cloud service providers, and AI research organizations requiring extreme memory capacity and compute density for generative AI model serving, scientific simulation, and high-performance computing. The system is supported by AMD's ROCm open software platform, providing compatibility with major AI frameworks including PyTorch and TensorFlow, and enabling straightforward workload migration from existing GPU infrastructure.

Ideal for

• Large language model inference and serving for models with hundreds of billions of parameters that benefit from in-memory residency across 192 GB per GPU
• Generative AI and multi-modal foundation model training requiring high aggregate GPU memory capacity and inter-GPU bandwidth
• High-performance computing and scientific simulation workloads demanding FP64 double-precision compute throughput at data center scale
• AI-driven drug discovery, protein structure prediction, and genomics workloads that require large memory footprints and fast matrix operations
• Enterprise AI infrastructure consolidation, reducing node count for large-model workloads through per-node memory density of 1.536 TB
• HPC cluster deployment in combination with HPE Cray Slingshot or InfiniBand fabric for tightly coupled parallel scientific computing

Technical specifications

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