Skip to content

AMD Instinct MI325X Accelerators

AI accelerators built on the AMD CDNA 3 architecture

Visit official site
Category AI models

What AMD Instinct MI325X Accelerators is

The hardware is described as AI accelerators based on AMD CDNA 3, aimed at training, fine-tuning and inference across systems and racks.

What you can do with it

  • Carry out foundation model training
  • Run a fine-tuning workload
  • Serve inference requests

Getting started

  • Confirm the system supports the card
  • Install the AMD software and drivers
  • Fit the card and do the base setup

Specifications

  • 6.0TB/s of bandwidth is listed
  • A 1.3 times bandwidth figure appears
  • Dell, HPE, Lenovo and Supermicro systems are named

Pros & cons

✓ What we like

  • FP16 and FP8 peak figures are given against H200
  • Mistral 7B FP16 inference performance is listed

! What to watch out for

  • Server or workstation compatibility must be checked
  • Tuning relies on AMD software tools

FAQ

Which architecture is used?

The material names AMD CDNA 3.

What memory is listed?

The feature list states 256GB HBM3E with 6.0TB/s bandwidth.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review