AMD Instinct MI300X Platform (8-GPU OAM baseboard)

The AMD Instinct MI300X is built on AMD's CDNA 3 compute architecture, integrating 8 GPU compute dies and 12 HBM3 memory stacks into a single Accelerator Processing Unit (APU) package using advanced 3D chiplet packaging. With 192GB of HBM3 memory per accelerator and 5.2TB/s of aggregate memory bandwidth, the MI300X is purpose-designed to hold extremely large AI models entirely within a single device, eliminating memory-constrained multi-GPU partitioning requirements that burden smaller accelerators. The OAM (Open Accelerator Module) baseboard variant houses 8 MI300X OAM modules in a shared high-bandwidth topology, delivering up to 1,536GB of total HBM3 capacity and 41.6 TB/s of aggregate memory bandwidth across the platform.

The MI300X platform leverages AMD's ROCm open software stack for GPU compute, providing compatibility with industry-standard frameworks including PyTorch, TensorFlow, JAX, and HIP. Native support for FP8, FP16, BF16, FP32, and INT8 data types allows flexible precision strategies for both training and inference, while the 383 TFLOPS of FP8 peak throughput per accelerator enables high-throughput token generation for large language models (LLMs) with billions to hundreds of billions of parameters. The platform integrates AMD Infinity Fabric interconnect technology for peer-to-peer GPU communication, supporting high-bandwidth, low-latency data exchange between all 8 accelerators on the baseboard without requiring host CPU involvement.

The 8-GPU OAM baseboard form factor is optimized for integration into hyperscale data center servers and AI-focused rack deployments. Each accelerator connects to the host system via PCIe Gen 5 x16, and the platform supports AMD's 7th Gen EPYC server processors as host CPUs, enabling a coherent memory fabric across CPU and GPU subsystems in supported configurations. The MI300X platform has achieved broad adoption among hyperscale cloud providers and AI-focused enterprises deploying large generative AI services, making it one of the most capacity-dense and bandwidth-rich accelerator platforms available for production AI infrastructure in 2024.

Ideal for

• Large language model (LLM) inference serving for models exceeding 70 billion parameters, where 192GB per accelerator enables single-device hosting of full model weights
• Distributed pre-training and fine-tuning of foundation models across multi-node clusters using AMD Infinity Fabric and RCCL collective communication libraries
• Generative AI application hosting including multimodal models, code generation systems, and retrieval-augmented generation (RAG) pipelines requiring high memory capacity
• High-performance computing (HPC) scientific simulation workloads such as molecular dynamics, climate modeling, and computational fluid dynamics leveraging FP64 double-precision compute
• AI recommendation system training and inference at hyperscale, processing large embedding tables that benefit from dense on-device HBM3 memory capacity
• Enterprise AI platform consolidation, reducing multi-node inference cluster footprint by fitting entire large model deployments onto fewer physical accelerator nodes

Technical specifications

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