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
Is it possible to play on AMD BC-160 and Instinct Mi50 mining graphics cards? It took us a week of testing, a trip to the service center, and the development of a unique ECHO to answer. 00:00 - MC intro. 02:30 - What is

AMD Instinct

AMD's data centre GPU line for deep learning and HPC

Photo: Мой Компьютер, CC BY 3.0, via Wikimedia Commons

AMD Instinct is AMD's brand of data centre GPUs, launched in 2017 as Radeon Instinct and shortened to Instinct before the MI100 arrived in November 2020. The line targets deep learning, neural networks and high-performance computing, and its accelerators power leading supercomputers such as Frontier.

Share on WhatsApp

  • Launched 20 June 2017 as Radeon Instinct, renamed Instinct before MI100
  • MI300X offers 304 compute units and 192 GB of HBM3
  • MI355X reaches up to 9.2 PetaFLOPS of FP4 compute
  • Frontier, using Instinct GPUs, topped TOP500 from June 2022
  • Transistor counts span 5.7 billion to 185 billion across the line

What it is

AMD Instinct is the name AMD gives to its data centre graphics processors. These are not cards for games or desktop displays; they are built to run deep learning, artificial neural network and general-purpose computing workloads on a GPU. The brand took over from the FirePro S line in 2016 and first shipped as Radeon Instinct on 20 June 2017. AMD shortened the name to Instinct before the MI100 launched in November 2020. The range sits against Nvidia's data centre GPUs and Intel's Xeon Phi and Data Center GPU products. Models carry the MI prefix and use between 36 and 304 compute units.

How it came about

Three Radeon Instinct cards were announced on 12 December 2016 and reached customers on 20 June 2017, each built on a different GPU design. The brand name was trimmed to AMD Instinct ahead of the MI100 in November 2020. In June 2022, machines pairing AMD Epyc CPUs with Instinct GPUs took the first four places on the Green500 ranking of power-efficient supercomputers, with a lead of more than 50% over the rest. Frontier, one of those systems, has been the fastest machine on the TOP500 list since June 2022 and through 2023.

Inside the chip

Early cards used Graphics Core Next designs, from GCN 3 through GCN 5. Later generations moved to the CDNA family, built purely for compute. The CDNA 1 generation dropped every rendering resource and added matrix processing units instead. CDNA 3 uses a chiplet layout: several separate dies are placed on one interposer and linked by AMD's Infinity Fabric, a high-speed on-package connection. The MI300A mixes 24 Zen 4 x86-64 CPU cores with four CDNA 3 GPU dies, while the MI300X swaps the CPU for more GPU logic. CDNA 4 adds FP4 and FP6 low-precision number formats alongside FP8 and FP16.

How it was made

Transistor counts across the line run from 5.7 billion on the Polaris 10 based MI6 to 185 billion on CDNA 4 parts. Process nodes improved from 28 nm and 14 nm on the earliest cards to 7 nm, 6 nm, 5 nm and finally TSMC's 3 nm N3 process for the MI350 series. MI300 and MI350 parts use TSMC packaging methods, CoWoS and InFO, to mount multiple chiplets on a shared interposer.

AMD Instinct
A drawing made by EveryChip from what this page says. It is a diagram, not a picture of the real chip.

Where it went

Instinct accelerators are aimed at inference and training of neural networks, plus high-performance computing. The MI6 and MI8 were positioned mainly for inference work. MI300X targets generative AI tasks such as natural language processing, computer vision and deep learning. Systems built from Epyc processors and Instinct cards populate the top of the Green500 efficiency list, and the Frontier supercomputer runs on this combination. Cards support PCIe 5.0 and CXL 2.0 links to the rest of a system.

Versions and family

The first three products were MI6, MI8 and MI25. MI50 and MI60, based on Vega 20, were the last to carry Radeon branding and the last able to drive a display. The MI100 series introduced CDNA 1. MI300A and MI300X use CDNA 3, and MI350X and MI355X use CDNA 4 with 288 GB of HBM3E. In 2026 AMD issued the MI350P as a PCIe card with roughly half the compute and power of the MI350X.

Why it matters

Instinct hardware underpinned a shift in supercomputing towards AMD parts. In June 2022, Epyc plus Instinct systems held the first four Green500 slots and led rivals by more than 50% on efficiency. Frontier, built on that pairing, has topped the TOP500 list since June 2022. The line gives AMD a direct answer to Nvidia and Intel in machine learning and GPGPU computing.

Problems and reception

MI6, MI8 and MI25 all support MxGPU, AMD's technology for sharing one GPU between several users. Software support runs through ROCm, with the MIOpen library handling deep learning and working with frameworks including TensorFlow, Caffe and PyTorch's predecessor Torch. MIOpen is meant to stand against the deep learning parts of Nvidia's CUDA stack.

Questions people ask about AMD Instinct

What is AMD Instinct for?

It is AMD's line of data centre GPUs. The cards accelerate deep learning, artificial neural network work and high-performance or general-purpose GPU computing, rather than driving games or desktop graphics. They compete with Nvidia's data centre GPUs and Intel's Xeon Phi and Data Center GPU products.

How does the MI300X differ from the MI300A?

The MI300A is an accelerated processing unit with 24 Zen 4 CPU cores alongside four CDNA 3 GPU dies, giving 228 compute units and 128 GB of HBM3. The MI300X removes those CPU cores in favour of extra GPU logic and memory, reaching 304 compute units and 192 GB of HBM3 for generative AI.

What memory do these accelerators use?

Early cards used GDDR5 or first-generation High Bandwidth Memory. Later parts moved to HBM2, HBM3 and, on the MI350 series, 288 GB of HBM3E with 8 TB/s of bandwidth. The MI300 family delivers 5.3 TB/s of memory bandwidth.

Which supercomputers use Instinct?

In June 2022, systems pairing AMD Epyc CPUs with Instinct GPUs held the top four places on the Green500 efficiency list. Frontier, one of them, has been the fastest machine on the TOP500 list since June 2022 and as of 2023.

What software runs on Instinct cards?

AMD's ROCm open compute stack covers the line. MIOpen, its deep learning library, works with frameworks such as TensorFlow, Caffe, MXNet and others, and code can be written in OpenCL or Python. MxGPU lets several users share one GPU.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "AMD Instinct", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

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.