AMD Instinct MI325X 256GB OAM Accelerator

The AMD Instinct MI325X is a high-memory AI accelerator built on AMD's CDNA3 compute architecture, delivering a substantial leap in memory capacity for large-scale AI and HPC workloads. At its core, the MI325X features 256GB of HBM3e memory — double the capacity of the MI300X — providing an expansive on-device memory pool that enables larger AI models, bigger batch sizes, and reduced reliance on multi-node memory offloading. The accelerator is packaged in the Open Accelerator Module (OAM) form factor, designed for integration into OAM-compliant server platforms and AI infrastructure at scale.

The MI325X is engineered for the demands of frontier generative AI inference and training, supporting mixed-precision compute across FP64, FP32, FP16, BF16, FP8, and INT8 data types. Its high-bandwidth memory subsystem, built on HBM3e technology, delivers significantly increased memory bandwidth compared to previous generations, directly addressing the memory-bound bottlenecks common in large language model (LLM) inference and transformer-based architectures. AMD Infinity Fabric interconnect technology enables high-speed peer-to-peer communication in multi-accelerator configurations.

As the newest addition to AMD's CDNA3 Instinct accelerator family, the MI325X is supported by the ROCm open software platform, providing compatibility with widely adopted AI frameworks including PyTorch and TensorFlow, as well as AMD's optimized libraries for math, deep learning, and communication primitives. Enterprises deploying the MI325X benefit from a fully open software ecosystem alongside the hardware's exceptional memory headroom, making it well-suited for organizations pushing the boundaries of model scale in production AI infrastructure.

Ideal for

• Large language model inference serving for frontier-scale models exceeding 100 billion parameters, where on-device memory capacity directly determines achievable context length and batch throughput
• Generative AI model training for transformer-based architectures requiring high aggregate memory bandwidth and capacity to hold model weights, optimizer states, and activations on a single accelerator
• Retrieval-augmented generation (RAG) pipelines and embedding workloads that benefit from large memory pools to cache extensive vector representations and model components simultaneously
• High-performance computing simulations in scientific research, including molecular dynamics, climate modeling, and computational fluid dynamics, leveraging FP64 throughput and large memory capacity
• Multi-modal AI workloads combining vision, language, and structured data models within a single inference session, where memory capacity is a primary architectural constraint
• Enterprise AI infrastructure consolidation, reducing the number of nodes required for large model deployment by maximizing per-accelerator memory and compute density in OAM-based server platforms

Technical specifications

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