AMD Instinct MI430X OAM

The AMD Instinct MI430X is the HPC and sovereign-AI member of the CDNA 5 based Instinct MI400 series, and a genuinely distinct SKU from the AI-focused MI455X rather than a configuration of it: AMD tailors each MI400-series part to a precision envelope, keeping full hardware FP64/FP32 execution on the MI430X while the MI440X and MI455X are optimised for FP4/FP8/BF16. AMD publishes a single compute figure for it — 288 TFLOPS of peak hardware-based FP64 (Vector, dense) — which AMD claims is roughly 8.7x the hardware FP64 peak of NVIDIA's Vera Rubin, whose double precision leans on emulation. It carries 432 GB of HBM4, the same capacity as the MI455X, and supports UALink for scale-up alongside AMD Infinity Fabric. AMD has published no FP32, FP16, BF16, FP8, FP4 or INT8 peak for this part, and no TDP, process node, transistor count or compute-unit count, so those fields are deliberately empty rather than borrowed from the MI455X. Memory bandwidth is left blank because AMD's stated "2.3 TB/s" cannot be right for a 432 GB HBM4 device and the credible alternatives (19.6 TB/s and 23.3 TB/s) conflict across AMD's own pages. Announced deployments are the US DOE/ORNL Discovery system (2028), Europe's Alice Recoque exascale system with GENCI/CEA, and the Herder system. Availability is expected in 2027.