AMD Instinct MI300A 128GB HBM3 APU Accelerator OAM

The AMD Instinct MI300A 128GB HBM3 APU Accelerator OAM (Gigabyte part 100-300000054) is built on AMD's groundbreaking Accelerated Processing Unit architecture, combining CPU and GPU compute dies alongside 128GB of unified HBM3 memory on a single OAM (OCP Accelerator Module) package. This heterogeneous design eliminates the PCIe bottleneck between host processor and accelerator by sharing a single, high-bandwidth memory pool across both CPU and GPU workloads, enabling dramatically lower data-movement latency and higher effective throughput for memory-bound HPC and AI inference tasks.

The MI300A is fabricated using a chiplet-based 3D stacking approach, integrating AMD CDNA 3 GPU compute dies with Zen 4 CPU cores and HBM3 stacks into one unified package. The CDNA 3 architecture delivers support for FP64, FP32, FP16, BF16, INT8, and FP8 precisions, making the accelerator suitable across the full spectrum of scientific simulation, large-scale model training, and high-throughput inference. The 128GB unified HBM3 memory subsystem provides substantial aggregate memory bandwidth, enabling large language models and numerical workloads that previously required multi-GPU configurations to run efficiently on a single OAM module.

As an OAM-form-factor device, the MI300A integrates directly into OCP-compliant UBB (Universal Baseboard) platforms and is targeted at hyperscale datacenter operators, national laboratories, and enterprises deploying AI supercomputing infrastructure. It is fully supported under the AMD ROCm open software ecosystem, providing compatibility with HIP, OpenCL, and popular ML frameworks including PyTorch and TensorFlow, as well as MPI and OpenMP for traditional HPC workloads.

Ideal for

• Large language model (LLM) training and fine-tuning where the unified 128GB HBM3 memory pool allows multi-billion parameter models to reside fully in-package without CPU-GPU data transfers
• High-performance computing (HPC) scientific simulations — including computational fluid dynamics, molecular dynamics, and climate modeling — that benefit from FP64 throughput and shared memory addressing
• AI inference serving at scale, leveraging FP8 and INT8 precision support to maximize token throughput per OAM module in rack-dense datacenter deployments
• Generative AI and foundation model development where large unified memory capacity reduces the need for model parallelism across multiple nodes
• Memory-bandwidth-intensive numerical workloads such as sparse linear algebra and graph analytics that saturate traditional discrete GPU memory hierarchies
• Converged HPC and AI pipelines where CPU preprocessing and GPU compute phases share data in a single unified address space, eliminating explicit data-staging overhead

Technical specifications

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