Back to basics: AMD's Instinct MI430X will be an FP64 accelerator

A sister model to the Instinct MI455X shifts the focus away from Artificial Intelligence, towards a classic accelerator for High-Performance Computing (HPC). AMD maximizes the GPU's computing power for double-precision floating-point calculations (FP64), which are necessary for accurate physical simulations, for example.

For the first time, AMD names a concrete value: the Instinct MI430X is expected to achieve more than 200 FP64 teraflops, i.e., over 200 trillion calculations per second. Simply extrapolated, 5000 GPUs would be sufficient for an exascale system. The previous Instinct top model MI355X is still a hybrid accelerator for both high-precision FP64 and AI calculations, achieving 78.6 FP64 teraflops.

Each Instinct MI430X carries 432 GByte of stacked memory in the form of HBM4, presumably divided into 12 modules. Together, they achieve a transfer rate of 19.6 TByte/s. Three supercomputers have already been announced that will use AMD's new Epyc Venice processors and MI430X GPUs: Herder in Germany, Alice Recoque in France, and Discovery in the USA.

For the Instinct MI430X, AMD must produce its own compute dies, whose processing units and data paths are designed for 64-bit values. The Instinct MI455X, with its AI focus, is intended for high computing power with lower precision, primarily 4 to 16 bits. When training and running AI models, precision plays a subordinate role because the results are only approximations anyway. Many fast calculations are more important here.

AI marketing even for HPC accelerators

However, AMD does not miss the opportunity to also bring in the topic of AI for the Instinct MI430X. According to this, systems with the GPU could deliver high-quality simulation data, which in turn would help AI models during training. Furthermore, this GPU also contains AI compute units for hybrid workloads, but fewer than the Instinct MI455X.

AMD has not yet commented on the availability of the Instinct MI430X. Shipping must begin by 2027 at the latest if the first supercomputers with the GPU are to be built. Initial deliveries in the second half of 2026 seem realistic.

Empfohlener redaktioneller Inhalt

Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.

Preisvergleiche immer laden Preisvergleich jetzt laden

Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.

(mma)