AMD recently announced the design direction of the MI430X, the first product of the Instinct MI400 series, targeting high-performance computing (HPC) and large-scale AI training environments, and simultaneously revealed the deployment plans for multiple supercomputers. The industry believes that AMD is accelerating the promotion of a new generation of accelerator platforms to form a new round of competition with NVIDIA's Rubin architecture.
AMD stated that MI430X adopts the next-generation CDNA architecture, which is a major update after MI300A and has significantly improved large-scale AI and HPC computing capabilities. Although the official specifications have not been announced, several core directions have been confirmed, including:
Supports HBM4 memory with a capacity of up to 432GB and a memory bandwidth of up to 19.6TB/s. Supports AI precision formats such as FP4 and FP8 and double-precision FP64 calculations required by HPC. Architectural improvements for hardware-intensive training and scientific computing.AMD pointed out that MI430X will be used in high-complexity scenarios such as large-scale model training, generative AI, energy materials research, and HPC simulation, and is expected to become an important foundation for the next wave of large-scale AI systems.
AMD emphasized that multiple top supercomputers have confirmed the introduction of MI430X, including the “Discovery” AI Factory system of the Oak Ridge National Laboratory in the United States, and the European Exascale platform “Alice Recoque”.
The two systems will be equipped with the next-generation EPYC "Venice" processor and built on the latest supercomputing platforms from HPE and Eviden. They will use the high bandwidth and high efficiency of MI430X to promote overall performance improvements in energy research, materials science, double-precision HPC and large-scale AI models. AMD said that these deployments reflect the strategic advantages of the MI400 series in the two fields of scientific computing and AI training, and have the conditions to support the construction of national-level AI capabilities.
AMD also announced that the higher-end Instinct MI455X is under development, which will compete with the NVIDIA Rubin series, focusing on large-scale model training performance, inference speed and energy efficiency. As AMD and NVIDIA simultaneously update their GPU/supercomputing platforms, the market expects that the AI accelerator market will usher in a new round of competition next year.
AMD Instinct MI430X: Powering the Next Wave of AI and Scientific Discovery AMD Unveils the Instinct MI430X AI Chip For “Leadership Performance” On AI Systems, Featuring HBM4 Memory & Next-Gen CDNA Architecture Further reading: TSMC 2nm blessing, Su Zifeng confirmed that AMD Instinct MI400 and Zen 6 architecture EPYC performance has greatly increased AMD adapts its own SoC to develop enhanced space-grade lidless organic packaging to strengthen its space-grade product lineup