Instinct MI250X 128GB

Specs, explained

Used for HPC and AI training

1. 01128 GB of HBM2E memory

   VRAM

   128 GB

   Memory type

   HBM2E

   Memory bandwidth

   3.2 TB/s

   HBM (High Bandwidth Memory) is stacked in layers and sits right next to the processor on the same package. That short, wide path is what gives accelerators with HBM their memory bandwidth, so they are usually chosen for work limited by how fast data reaches the cores, such as large-model inference and HPC simulation.

   VRAM is the working space on the card. Model weights, activations and datasets have to fit in it, or be split across several GPUs. Rule of thumb for the weights alone at 16-bit precision (about 2 GB per billion parameters): up to roughly 64 billion parameters, before activations and cache. This is arithmetic on the recorded VRAM, not a benchmark.

2. 02CDNA 2 architecture

   Architecture

   CDNA 2

   Generation

   Instinct MI200

   Launch year

   2021

   Lifecycle status

   Legacy

   CDNA is AMD's compute-focused architecture for Instinct accelerators, separate from its RDNA graphics line. (General context about the architecture, not a statement about this specific card.)

   Launched in 2021. Kovara's records list it as a legacy part.

3. 03Form factor and power

   Form factor

   OAM Module

   TDP

   560 W

   Accelerator modules are mounted on a baseboard rather than slotted in as cards, the layout used for dense multi-accelerator servers.

   TDP is the rated thermal design power: the planning figure for power delivery and cooling in a server.

4. 04What it is used for

   Primary use (recorded)

   HPC / AI training

   These are the uses Kovara has recorded for this part. The notes explain the workload in general terms; they are not performance claims.

   • HPCScientific and engineering simulation, where numerical precision and memory bandwidth matter.
   • AI trainingTeaching a model from data. Usually the most compute- and memory-hungry phase, often spread across many GPUs.

Figures are Kovara master data, last verified 03 Sept 2026. Explanations describe the technology in general; they are not benchmarks and say nothing about price or availability.