AMD expands Instinct MI400 accelerator portfolio

AMD expands Instinct MI400 accelerator portfolio

AMD has expanded its MI400 accelerator configurations for AI workloads. The new MI450X provides 288 GB of HBM4, while updated MI430X specifications address scientific computing and MI455X testing provides further inference performance results.


IN Brief:

  • AMD has introduced the Instinct MI450X with 288 GB of HBM4 for its Helios rack architecture.
  • The MI430X targets scientific computing with native FP64 processing, 432 GB of HBM4, and memory bandwidth up to 23.3 TB/s.
  • AMD reports 21,333 tokens per second per MI455X GPU in a company-tested DeepSeek-R1 MLPerf Offline configuration.

AMD has expanded its Instinct MI400 accelerator portfolio with a new memory configuration, additional scientific computing specifications, and measured inference results for its Helios rack-scale computing architecture.

The developments, detailed on 12 November, include the introduction of the Instinct MI450X, further information about the MI430X, and new performance measurements for the MI455X. Each addresses different combinations of memory capacity, numerical precision, and accelerated computing requirements.

The MI450X provides 288 GB of HBM4 memory and is designed for integration within AMD Helios systems. It shares the core computing capabilities, memory bandwidth, and scale-up and scale-out connectivity characteristics of the MI455X, which provides a larger 432 GB memory configuration.

Both accelerators use AMD’s fifth-generation CDNA architecture. The MI450X includes 256 work group processors and supports theoretical peak HBM4 bandwidth of 23.3 TB/s, along with high-speed connections to other accelerators and the surrounding computing infrastructure.

Memory capacity is an important consideration when selecting hardware for large AI workloads. A model’s parameters must be stored and accessed during execution, while intermediate calculations and other operating data consume additional memory.

The requirement varies considerably between applications. Larger models may need more memory to store their parameters, while long input contexts and multiple simultaneous inference requests can increase the amount required during operation.

Where a workload fits within 288 GB, the MI450X provides an alternative to installing a higher-capacity accelerator. Applications requiring more memory may benefit from the MI455X, although memory capacity alone does not establish processing throughput or response time.

Performance also depends on how efficiently data can move between the memory subsystem and computing units. A processor with substantial arithmetic capability may spend time waiting for information if the workload cannot be supplied at a sufficient rate.

High-bandwidth memory addresses part of this constraint by placing multiple memory dies close to the accelerator and connecting them through a dense interface. HBM4 increases the available transfer capability, but actual application performance remains dependent on memory access patterns and the efficiency of the surrounding software.

The MI450X and MI455X use the Helios architecture and AMD ROCm software environment. This provides a common foundation for deploying compatible applications across the two memory configurations, although system integration and performance tuning remain necessary for individual workloads.

The MI430X is aimed more specifically at high-performance scientific computing. AMD has detailed a configuration combining native double-precision floating-point processing with 432 GB of HBM4 memory and bandwidth reaching 23.3 TB/s.

Double-precision computation, generally referred to as FP64, is important in numerical applications where calculations require greater precision than lower-precision formats can provide. Scientific simulation, computational fluid dynamics, and mathematical modelling can involve repeated calculations in which rounding errors accumulate.

The accuracy required depends on the numerical method and the characteristics of the problem being solved. Some algorithms can use reduced precision for selected calculations, while others rely heavily on native double-precision operations to maintain the required numerical behaviour.

AMD has examined 47 workload configurations across 11 scientific applications and benchmarks to investigate differences between native FP64 processing and methods that emulate double precision through other arithmetic operations.

The company has positioned the MI430X around these requirements rather than treating scientific computing as an extension of lower-precision AI acceleration. The distinction is important because peak performance figures quoted for AI inference do not establish equivalent performance in scientific applications.

The MI455X results address AI inference rather than scientific simulation. AMD reports throughput of 21,333 tokens per second per GPU when running DeepSeek-R1 on an AMD Helios configuration using an MLPerf Inference Offline test methodology.

The company compares that measurement with published NVIDIA Vera Rubin results and reports up to 30% greater throughput under the selected conditions. AMD conducted its own testing in October 2026 and compared the result with published MLPerf Inference v6.1 entries retrieved on 11 November.

The results have specific limitations. MLPerf Offline testing measures the processing throughput of a defined workload without reproducing every latency constraint associated with an interactive service. Performance may change with model selection, batch size, context length, software version, and the configuration of the complete system.

The published comparison also draws on measurements from different test submissions rather than a single independently controlled comparison of all competing hardware. It therefore supports the stated benchmark result but cannot establish a general performance advantage across every inference application.

AMD also cites the Helios platform’s memory bandwidth and accelerator connectivity as contributors to its performance. Interconnect capacity becomes important when computing workloads are distributed across multiple GPUs, because data must move between devices as well as between each accelerator and its local memory.

The completed system must coordinate accelerator communication, host processing, power delivery, cooling, and software execution. Bottlenecks in any of these areas can prevent installed arithmetic capacity from being used efficiently.

Direct liquid cooling forms part of AMD’s MI400 rack-scale hardware design. Managing heat is particularly important in dense accelerator installations, where high electrical power consumption and tightly packed components create substantial thermal loads.

The latest MI400 developments provide additional configuration choices for AI and scientific computing. The MI450X offers a smaller memory option within the same Helios architecture, the MI430X addresses native double-precision requirements, and the MI455X results supply a new performance measurement under defined inference conditions.

Further application testing will determine how those capabilities translate into sustained throughput, numerical performance, and operating costs across production workloads. The published specifications establish the hardware characteristics, while the benchmark results describe performance under the particular conditions reported by AMD.


Stories for you


  • Murata integrates processing into industrial motion sensors

    Murata integrates processing into industrial motion sensors

    Murata has introduced two integrated inertial sensors for industrial applications. The SCI36LT-K01 performs orientation calculations using embedded sensor fusion, while the K02 allows customer software to run on its Cortex-M4 processor.


  • LITEON details 72 kW AI rack power shelf

    LITEON details 72 kW AI rack power shelf

    LITEON has detailed additional power hardware for dense AI racks. Its electronica portfolio includes a 72 kW, 1U shelf, established 110 kW equipment, capacitor backup, and 800 V DC distribution concepts.