AMD expanded its Instinct MI400 Series GPU lineup with hardware updates tailored for scientific computing and artificial intelligence, featuring the newly introduced MI450X and fresh benchmark results for the MI455X, as reported by HPCwire.
The hardware additions target customers utilizing the Helios rack-scale system for high-performance computing, scientific simulation, and interactive AI inference workloads.
The company positioned the new hardware options around the Helios rack-scale architecture and the ROCm software stack. HPCwire reported that the MI430X GPU is engineered specifically for research workloads that combine traditional high-precision simulation with machine learning models.
“With advances in high-precision compute, new milestones in AI inference performance and an expanded range of configurations, we’re giving customers more options to power their next generation of AI and HPC workloads,” said Andrew Dieckmann, Corporate Vice President and General Manager, Data Centre GPU Business, AMD.
MI430X features native FP64 support and high HBM4 bandwidth
The MI430X features native double-precision FP64 support for scientific calculations, 432GB of HBM4 memory, and memory bandwidth reaching up to 23.3 terabytes per second, according to the official AMD announcement. These specifications address the demands of large simulations and massive scientific datasets where double-precision arithmetic remains mandatory.
Alongside hardware releases, AMD published research reviewing 47 workload configurations across 11 applications and benchmarks to evaluate when FP64 emulation suffices and when native double precision is necessary. The findings support continued demand for infrastructure that handles both artificial intelligence tasks and conventional high-performance computing jobs simultaneously.

Tom’s Hardware said that AMD’s Scott Tease explained that most supercomputing sites recognize AI will never replace traditional high-performance computing. Tease told Tom’s Hardware that building a heterogeneous compute stack remains the best path forward for bleeding-edge research.
MI455X on Helios delivers high throughput for DeepSeek-R1
In benchmark disclosures, AMD published measured inference figures for the MI455X running on the Helios system. The company reported that the setup delivered 21,333 tokens per second per GPU on DeepSeek-R1 during an MLPerf Inference Offline configuration test conducted in October, as detailed by iTWire.
According to AMD testing cited by HPCwire, that output represents throughput up to 30% higher than Nvidia’s latest published MLPerf Offline result for the Vera Rubin platform. iTWire reported that Helios provides up to 20% higher HBM bandwidth than Nvidia Vera Rubin alongside superior scale-up bandwidth.
Complementing the flagship part, AMD introduced the MI450X to provide Helios customers with a lower-memory configuration while retaining core compute, memory bandwidth, and connectivity. The MI450X equips 288GB of HBM4 memory per GPU, intended for workloads that operate efficiently within a smaller memory footprint.

Common questions about AMD Instinct MI450X and MI455X
What is the memory capacity of the new AMD Instinct MI450X?
The MI450X comes equipped with 288GB of HBM4 memory per GPU, offering a lower-memory alternative to the 432GB available on the MI455X model.
How does the MI455X perform in benchmark tests against competing hardware?
AMD reported that the MI455X delivered 21,333 tokens per second per GPU on DeepSeek-R1 in an MLPerf Inference Offline configuration, achieving throughput up to 30% higher than Nvidia’s latest published Vera Rubin offline result.
Which supercomputer uses AMD Instinct accelerators in Australia?
Australia’s most powerful research supercomputer, Pawsey’s Setonix system in Perth, runs AMD Instinct MI250X GPUs per its Top500 listing, as reported by iTWire.
What is the role of the MI430X in scientific research?
The MI430X is designed for research workloads combining traditional simulation with AI, featuring native FP64 support, 432GB of HBM4 memory, and memory bandwidth up to 23.3 terabytes per second.
Related reading