AMD Instinct MI430X Promises to Outperform NVIDIA Rubin: GPU Up to 6 Times Faster in a Specific Domain

AMD has provided a first glimpse of the capabilities of the Instinct MI430X, an accelerator intended for the HPC segment that will stand out sharply from the current market trend that is almost exclusively focused on artificial intelligence. The new solution will indeed be designed to prioritize double-precision computing (FP64), which remains fundamental in numerous scientific, simulation, and engineering fields.

In recent years, the evolution of high-end GPUs has mainly concentrated on AI workloads, focusing on reduced precision numerical formats such as FP8, FP6, and FP4 to enhance throughput and energy efficiency. This approach is ideal for training and inferring neural models but does not represent the optimal solution for traditional HPC applications, where numerical accuracy and stability of results remain paramount.

For this reason, AMD has decided to further differentiate the future Instinct MI400 family. Alongside the MI450X and MI455X models, primarily aimed at AI, there will be the MI430X, a solution specifically dedicated to classical supercomputing. The chip was officially confirmed last year, but now concrete initial indications about its performance have emerged.

According to the company’s statements, the Instinct MI430X will be capable of exceeding 200 TFLOPS native FP64, a value that would represent the highest level ever achieved by an HPC GPU. The card will also integrate 432 GB of HBM4 memory and will leverage an evolution of the CDNA architecture developed with advanced packaging and manufacturing technologies.

AMD claims that the new GPU will offer up to six times the FP64 performance of the future NVIDIA Rubin. However, the comparison requires some clarifications: Rubin is primarily designed as an AI accelerator and concentrates much of its capabilities on low-precision formats. NVIDIA indicates around 33 TFLOPS of native FP64 vector processing, with higher values achievable through Tensor Core and emulation algorithms. The MI430X, on the other hand, would be expressly designed for native FP64 computation.

The performance leap is also significant compared to current AMD offerings. The Instinct MI355X, MI325X, and MI300X are capped at around 80 TFLOPS of FP64, indicating that with the MI430X, the company intends to introduce substantial acceleration in traditional HPC applications.

AMD has also emphasized that the MI430X will not sacrifice low-precision AI capabilities, thus trying to combine HPC performance and AI acceleration within the same platform. However, it remains evident that the primary focus of the product is directed towards large scientific systems and exascale supercomputers.

Among the first confirmed projects is Discovery, the new supercomputer from the Oak Ridge National Laboratory in the United States, expected in 2028. The system will use Instinct MI430X accelerators alongside next-generation EPYC CPUs and will be employed by the U.S. Department of Energy for energy research, biology, national security, advanced materials, and manufacturing.

Europe will also have access to the new AMD platform through the Alice Recoque project, a French system destined to become one of the continent's major exascale supercomputers. According to preliminary information, the system is expected to surpass one exaFLOP of FP64 power.

The final specifications of the Instinct MI400 family, including those of the MI430X, should be officially presented during the AMD Advancing AI 2026 event scheduled for July 22 and 23.