AMD unveils the Instinct MI430X GPU, designed for high-precision compute and AI, featuring 288 TFLOPS FP64, 432GB HBM4, and CDNA 5 architecture.

AMD has introduced the Instinct MI430X GPU, marking a significant advancement in high-performance computing (HPC) and artificial intelligence (AI). Built on the AMD CDNA™ 5 architecture, the MI430X is engineered to address the growing demand for the convergence of high-precision compute and AI-driven workloads in scientific research and enterprise applications.

The AMD Instinct MI430X GPU delivers up to 288 TFLOPS of peak native FP64 performance and 9.2 PF of peak FP4 performance. It features 432 GB of HBM4 memory with a bandwidth of 23.3 TB/s, offering a 50% increase in memory capacity compared to Nvidia's Vera Rubin. This configuration enables larger datasets, such as meshes, atoms, and sequences, to remain resident on the GPU, reducing the need for data transfers across nodes and networks.

For multi-GPU configurations, the MI430X integrates seamlessly with a 6th Gen AMD EPYC™ Server CPU SP7 host, connecting four GPUs in a glue-less mesh using AMD Infinity Fabric™ technology. This design provides up to 384 GB/s unidirectional bandwidth per GPU and 12.3 Tbps of scale-up bandwidth across the node, optimizing performance for workloads dominated by halo exchanges and small-message collectives.

Compared to its predecessor, the AMD Instinct MI300A, the MI430X demonstrates significant generational improvements across various benchmarks:

The MI430X emphasizes native FP64 performance, addressing limitations of emulated FP64 techniques like Ozaki GEMM. According to AMD researchers, emulation benefits only a narrow subset of HPC workloads, such as dense GEMM operations, which constitute less than 10% of the HPC ecosystem.

Source link