AMD unveils Instinct MI430X GPU to power next-gen supercomputers

AMD has introduced the Instinct MI430X GPU, a processor designed to address the requirements of the most demanding artificial intelligence and high-performance computing (HPC) workloads in research and enterprise. The GPU is equipped with 432GB of HBM4 memory and a 19.6TB/s memory bandwidth, aimed at accelerating AI model training and a wide range of complex scientific simulations.

Supercomputing deployments

The MI430X will be deployed in forthcoming supercomputers including Discovery at Oak Ridge National Labouratory in the United States and Alice Recoque in Europe. These installations follow a succession of systems relying on AMD chips, such as Frontier and El Capitan-recognised as some of the fastest supercomputers in the world. AMD technologies are also present in ENI's HPC6, the largest commercial supercomputer, used primarily for energy sector research.

Architectural advances

The launch of the MI430X marks the next stage in AMD's high-performance computing roadmap, which has seen three generations of GPU developments. The MI250X powered Frontier with a dual-GPU design and a high-speed link to CPUs, while MI300A in El Capitan combined CPU and GPU cores into a single unit, improving integration and efficiency. The MI430X now supports FP64 precision for scientific computation and offers enhanced AI capabilities, enabling both tasks to run on the same platform without performance loss.

System integration

Discovery, at Oak Ridge, combines the new MI430X GPUs with AMD's next-generation EPYC Venice CPUs on an HPE Cray GX5000 platform. This setup is designed to support large-scale AI model development and high-end scientific workloads, including in energy and generative AI research. The Alice Recoque system in Europe adopts a similar approach, integrating MI430X GPUs and next-generation EPYC CPUs with Eviden's BullSequana XH3500 platform to achieve high performance in double-precision HPC and AI tasks. Both systems are engineered to address stringent energy efficiency targets and offer extensive memory bandwidth for scientific computing.

Software and performance

The MI430X works with AMD's ROCm software stack, designed for compatibility with industry-standard machine learning frameworks including PyTorch, TensorFlow, and JAX. This compatibility allows researchers and institutions using the GPUs to scale their applications across entire data centres or supercomputing environments. The GPU's combination of FP4, FP8, and FP64 precision support matches the needs of both AI research and computational science, minimising memory bottlenecks and supporting large models and data sets.

National and enterprise focus

AMD describes the Instinct MI430X GPU's deployment as part of an ongoing move towards sovereign AI infrastructure. By providing open hardware and software platforms, the company aims to support governments and commercial organisations looking to develop and secure their own AI and HPC capabilities.

"From Discovery in the U.S. to Alice Recoque in Europe, the AMD Instinct MI430X is accelerating a new generation of AI factories and scientific platforms built on open standards and energy-efficient performance," said AMD.