AMD's MI430X Packs 432GB of HBM4 to Take On Nvidia's Vera Rubin

AMD unveiled the Instinct MI430X GPU on Thursday, a chip built on its CDNA 5 architecture and aimed squarely at the spot where high-precision compute meets AI. The company is pitching it at scientific research and enterprise workloads that need both. On paper it hits 288 TFLOPS of peak native FP64 and 9.2 PF of peak FP4 performance.

The memory spec is the headline grabber: 432GB of HBM4 running at 23.3 TB/s of bandwidth. AMD says that's 50% more memory capacity than Nvidia's Vera Rubin. The pitch is that big meshes, atoms and sequences can stay resident on the GPU instead of being shuttled across nodes and networks.

For multi-GPU rigs, the MI430X pairs with a 6th Gen AMD EPYC Server CPU SP7 host and links four GPUs in a glue-less mesh over AMD Infinity Fabric. AMD lists up to 384 GB/s of unidirectional bandwidth per GPU and 12.3 Tbps of scale-up bandwidth across the node. That's aimed at workloads dominated by halo exchanges and small-message collectives. AMD also says the MI430X shows generational gains over its predecessor, the Instinct MI300A, across a range of benchmarks.

AMD is making a blunt case against FP64 emulation tricks like Ozaki GEMM. Its researchers say emulation only helps a narrow slice of HPC work — dense GEMM operations, which they peg at less than 10% of the ecosystem. Native FP64, in AMD's telling, sidesteps the extra memory footprint and performance penalties that come with faking it.

The target list spans climate modeling, drug discovery, advanced manufacturing and national security. AMD's bet is that researchers and engineers want one part that handles high-precision math and AI together rather than splitting the job across two.