AMD Instinct™ MI400 Series GPUs

AMD Instinct™ MI455X GPUs power AMD Helios rackscale solution and provide the leadership per-device capability that aggregates through repeatable 4-GPU trays into the 72-GPU AMD Helios to realize predictable, serviceable exaFLOP-class results.

AMD Instinct™ MI455X GPUs provide:

• Up to 40 PFLOPs peak 4-bit performance
• Up to 20 PFLOPs peak 8-bit performance
• 432 GB HBM4
• 23.3 TB/s memory bandwidth

These capabilities aggregate through repeatable 4-GPU trays into the 72-GPU AMD Helios rackscale solution.

Memory has become one of the most important considerations for AI infrastructure as models continue to increase in size and complexity. Each AMD Instinct MI455X GPU provides 432 GB of HBM4 memory, while a full Helios rack delivers 31 TB of HBM4 memory. Combined with high memory bandwidth, this enables customers to support larger models, reduce infrastructure complexity and efficiently scale memory-intensive inference and training workloads.

AMD Instinct™ MI430X GPU is a new GPU purpose-built for sovereign AI, scientific computing and high-performance computing (HPC). AMD Instinct MI430X GPU is designed to support converged AI and HPC workloads by combining industry-leading HBM4 memory, high-performance FP64 compute and the open AMD ROCm™ software platform. By offering 288 TFLOPS peak hardware-based FP64 performance, AMD Instinct MI430X is built to help governments, research institutions and national laboratories tackle the world's most demanding simulations and AI applications.

Scientific computing and sovereign AI increasingly require infrastructure capable of supporting both traditional HPC and modern AI. AMD Instinct MI430X enables organizations to consolidate these workloads on a single platform while improving efficiency, simplifying infrastructure and maintaining the openness and flexibility required for national-scale deployments.

 For AI gigafactories and high-performance computing centers, AMD Instinct MI430X is designed to deliver the numerical fidelity and throughput required for the most demanding workloads. As next-generation AI models increasingly rely on training data generated from high-fidelity simulations, AMD Instinct MI430X helps researchers and institutions build AI systems on accurate scientific data – enabling greater confidence in AI-driven discovery across fields such as climate science, materials research, nuclear engineering and fluid dynamics.

AMD Instinct™ MI430X GPUs deliver uncompromised accuracy and throughput across converged HPC and AI workflows, with 288 TFLOPS hardware-based FP64 for scientific computing alongside support for modern AI data types and software capabilities. With industry-leading 432GB of HBM4 and 2.3 TB/s memory bandwidth, AMD Instinct MI430X is built for large simulations and memory-intensive models – reducing server count, lowering infrastructure complexity, and improving efficiency when scaling massive datasets.

Yes, the next US DOE/ORNL Discovery system and two European systems the Alice Recoque and Herder systems are expected to be deployed with AMD Instinct MI430X GPUs.  

Powered by AMD ROCm™ software, AMD Instinct MI430X provides an open foundation with programming models, compilers, libraries, runtimes, and deployment tools for both HPC and AI. This commitment to open standards supports interoperability and long-term portability across software, system, and networking choices – helping sovereign operators innovate without lock-in.

The AMD Instinct MI430X is estimated to have 8.7X better hardware-based FP64 peak FLOP performance over Nvidia Vera Rubin.

ENDNOTE: AMD Instinct MI400-001A - Based on engineering projections by AMD performance labs in July 2026, to estimate peak theoretical precision of the AMD Instinct™ MI430X GPU using FP64 (Vetor) datatype vs. AMD Instinct MI300X GPU vs. Nvidia Rubin GPU. Results subject to change when products are released in market.

The AMD Instinct MI430X GPUs are expected to be available in 2027. More details will be provided in the future.