AMD MI430X targets 200 TFLOPs of FP64 performance for future supercomputers

AMD has previewed its Instinct MI430X accelerator, a new HPC focused GPU designed to deliver very high FP64 performance for scientific and technical workloads.

The MI430X is part of AMD’s upcoming MI400 series. While the MI450X is aimed mainly at AI acceleration, the MI430X is built for classic high performance computing tasks where FP64 precision still matters.

AI hardware often focuses on lower precision formats such as FP4, FP6, and FP8 because they are useful for neural networks. HPC workloads are different. Fields such as physics, energy research, biology, national security, advanced materials, and large scale simulations often still need high precision compute. That is where AMD is positioning the MI430X.

AMD says the MI430X can reach up to 200 TFLOPs of native FP64 vector performance. That would make it one of the strongest FP64 GPUs announced so far. The company also claims this gives it a major advantage over NVIDIA’s upcoming Rubin GPU in classic HPC workloads.

According to the comparison shared, NVIDIA Rubin offers 33 TFLOPs of native FP64 vector performance, while AMD’s MI430X reaches 200 TFLOPs. Rubin can reach 200 TFLOPs using Tensor Core based FP64 matrix emulation, but AMD’s claim focuses on native FP64 vector compute.

Here is the basic comparison:

The MI430X uses AMD’s next generation CDNA architecture and HBM4 memory. AMD says the chip will also offer strong low precision AI capabilities, so it is not limited to traditional HPC alone. The goal is to serve systems that need both scientific compute and modern AI performance in one accelerator.

The chip is already tied to major supercomputer plans. Oak Ridge National Laboratory will use AMD Instinct MI430X accelerators and AMD EPYC CPUs in its Discovery supercomputer, planned for deployment in 2028. The system is expected to support work across energy, biology, national security, advanced materials, and manufacturing.

In Europe, the MI430X will also be used in the Alice Recoque system with next generation EPYC CPUs. That system is planned as an exascale class supercomputer for European research.

The MI430X shows that AMD still sees high precision HPC as a major market, even as most attention has shifted toward AI accelerators. For data centers running scientific simulations and AI workloads together, AMD wants the MI430X to be a single accelerator that can handle both sides of the job.