AMD Announces Instinct MI430X GPU with Native FP64 and HBM4 Memory

According to AMD Blogs:

Hardware Architecture and Specifications

AMD reports that the Instinct MI430X GPU is designed for the convergence of high-precision scientific computing and AI workloads. Built on the CDNA 5 architecture, it delivers up to 288 TFLOPS of peak native FP64 performance alongside 432 GB of HBM4 memory running at 23.3 TB/s bandwidth. For multi-GPU scaling, a node pairs a 6th Gen EPYC server CPU with four GPUs using a fully connected, switchless Infinity Fabric mesh.

Research on FP64 Emulation Limits

In a paper presented at SC26, AMD researchers evaluated the limits of Ozaki GEMM emulation on HPC systems. Their analysis suggests that less than 10% of the broader HPC ecosystem features workloads dense and GEMM-shaped enough to benefit from emulation, indicating that native double precision remains essential for stencils, sparse linear algebra, and molecular dynamics.

System Deployments and Projections

AMD states that the MI430X has been selected for several prominent European and US scientific systems, including Discovery at Oak Ridge National Laboratory, Alice Recoque in France, and Herder in Germany. Performance figures provided by AMD are projected estimates based on engineering benchmarks comparing the MI430X to the previous-generation MI300A APU.