News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
AMD Instinct MI325X

AMD Instinct MI325X

AI accelerator with 256GB HBM3E memory

The AMD Instinct MI325X is an artificial intelligence graphics processor launched by AMD. It features 256 gigabytes of high-bandwidth memory and uses the OAM module format for server deployment.

Share on WhatsApp

  • Features 256GB of HBM3E high-bandwidth memory
  • Uses Open Accelerator Module package format
  • Supports PyTorch and TensorFlow via ROCm stack
  • Succeeds the AMD Instinct MI300X accelerator
  • Designed for eight-GPU Universal Baseboard systems

What it is

The AMD Instinct MI325X is a specialised graphics processing unit designed for artificial intelligence workloads. AMD developed this chip to compete in the data centre market, focusing on large language model inference and training. The device stands out due to its substantial onboard memory capacity, which allows it to handle larger AI models than some competing hardware. It serves as an incremental upgrade over the previous MI300X generation, offering improved memory bandwidth and tuned core frequencies for better performance in server environments.

How it came about

AMD introduced the MI325X during an event that also featured its EPYC 9005 central processors. The launch positioned the chip as a direct competitor to the NVIDIA H200 accelerator. This release followed earlier previews and marked a step in AMD's strategy to capture more share in the AI hardware sector. The company has outlined a roadmap that includes the MI350 series for mid-2025 and the MI400 series for 2026, indicating a continuous development cycle for its Instinct product line.

Inside the chip

The processor utilises the CDNA 3 architecture, which AMD optimised for AI tasks. A key design feature is the integration of 256 gigabytes of HBM3E memory. HBM stands for High Bandwidth Memory, a type of dynamic random-access memory that provides faster data transfer rates than standard memory. This large memory pool enables the chip to store bigger neural networks locally. The GPU cores operate at tuned frequencies to maximise throughput. The design prioritises memory capacity and bandwidth to support complex inference tasks without needing to swap data to slower storage.

How it was made

The chip is packaged in an Open Accelerator Module form factor. This standardised module includes an air-cooled heatsink and is designed for high-density server installations. The physical label on the module identifies it as the MI326X IFX OAM. These modules are not intended for standalone use but are built to fit into Universal Baseboard platforms. A single baseboard can hold eight of these accelerators, allowing for scalable computing power within a single server chassis.

Where it went

Data centres deploy the MI325X for deep learning applications. The hardware supports popular software frameworks such as PyTorch, TensorFlow, and JAX through the ROCm stack. ROCm is AMD's open-source software platform for GPU computing. The chip is particularly suited for tasks that require large amounts of video memory, such as running large language models. It targets enterprise customers who need to perform AI inference at scale, where keeping the entire model in fast memory reduces latency and improves efficiency compared to systems with less memory.

Versions and family

The MI325X succeeds the MI300X and precedes the upcoming MI355X. AMD plans to release the MI350 series in 2025, which will move to a three-nanometre process and support new data formats like FP4 and FP6. The subsequent MI400 series, expected in 2026, will introduce a major architectural overhaul. The current MI325X represents an intermediate step, refining the existing CDNA 3 design with enhanced memory rather than a complete architectural change. This approach allows AMD to offer immediate improvements while preparing for more significant leaps in future generations.

Why it matters

This accelerator highlights AMD's focus on memory capacity as a competitive advantage. By offering significantly more memory than rival products, AMD aims to attract customers who struggle with model size limitations. The chip helps establish AMD as a viable alternative to dominant market players in the AI sector. Its adoption depends on the maturity of the ROCm software ecosystem, which continues to expand support for various AI libraries. The device contributes to the broader industry trend of increasing memory bandwidth to keep pace with growing AI model complexities.

Problems and reception

Industry observers note that the MI325X provides a compelling option for users constrained by memory limits. Comparisons with the NVIDIA H200 emphasise the 256 gigabyte capacity versus the competitor's 141 gigabytes. Some users report that information about actual usage and software support can be fragmented. The requirement to use Linux for PyTorch support via ROCm is a noted constraint for some potential adopters. Despite these software hurdles, the hardware is recognised for its raw specifications and potential cost benefits in specific deployment scenarios.

Questions people ask about AMD Instinct MI325X

How much memory does the MI325X have?

The AMD Instinct MI325X contains 256 gigabytes of HBM3E memory. This is significantly more than the 141 gigabytes found in the competing NVIDIA H200, allowing for larger AI models to run on a single GPU.

What software frameworks does it support?

The chip supports deep learning frameworks including PyTorch, TensorFlow, and JAX. This support is provided through AMD's ROCm software stack, which currently runs on Linux operating systems.

How is the MI325X packaged for servers?

It uses the Open Accelerator Module form factor with an air-cooled heatsink. These modules are designed to be installed in groups of eight on a Universal Baseboard platform for high-density computing.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on AMD Instinct MI325X yet, so the facts on this page come from the sources listed above: tomshardware.com/tech-industry/artificial-intelligence/amd-announces-m, servethehome.com/amd-instinct-mi325x-launched-and-the-mi355x-is-coming. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    Photo credits on this page

    The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.