CDNA
AMD microarchitecture for data centre GPU accelerators
CDNA is a graphics processing unit microarchitecture designed by AMD for high-performance computing in data centres. It powers the Instinct accelerator line, evolving from monolithic dies to complex multi-chip modules across three generations.
- Designed by AMD for data centre compute workloads
- First generation launched in November 2020 with MI100
- Uses multi-chip modules in second and third generations
- Manufactured by TSMC on 7 nm, 6 nm and 5 nm nodes
- Supports up to 146 billion transistors in CDNA 3
What it is
CDNA, short for Compute DNA, is a microarchitecture for graphics processing units built specifically for data centre workloads. AMD designed it to handle heavy computational tasks rather than consumer gaming graphics. It serves as the foundation for the AMD Instinct series of accelerator cards. The architecture removes traditional graphics hardware like display engines to focus entirely on compute performance. It supports advanced data types and matrix operations essential for artificial intelligence and scientific simulation.
How it came about
AMD announced the first generation in March 2020, with the MI100 card launching later that year. This initial version used a single large die made on a seven-nanometre process. The second generation arrived in late 2021, introducing a multi-chip module design that connected two dies on one package. AMD continued this trend with the third generation in 2023, which combined multiple chiplets using advanced three-dimensional packaging techniques to create massive accelerator complexes.
Inside the chip
The architecture organises compute units into asynchronous engines that manage command execution independently. Each unit contains vector processors that handle wavefronts of threads over several cycles. Dedicated matrix hardware accelerates specific mathematical operations used in machine learning. Memory systems use high-bandwidth stacks connected via wide buses. Later versions added support for double-precision floating-point calculations at full speed and improved cache bandwidth to feed data to the processing cores more efficiently.
How it was made
TSMC manufactures the chips using various process nodes, starting with N7 for the first generation. The second generation moved to the N6 node, while the third generation utilises both five-nanometre and six-nanometre processes for different chiplets. The most complex packages contain billions of transistors spread across fifteen unique dies. Advanced packaging methods stack these components vertically, connecting them with high-speed links to act as a single coherent processor.
Where it went
These processors power the AMD Instinct line of data centre graphics cards. They are deployed in high-performance computing clusters and artificial intelligence training systems. The cards fit into standard server slots or specialised accelerator modules defined by the Open Compute Project. Some variants integrate central processor cores to function as accelerated processing units, allowing them to run general-purpose code alongside heavy parallel workloads in supercomputers.
Versions and family
The first generation, known as Arcturus, featured a single monolithic die. The second generation, Aldebaran, introduced dual-die packages and a monolithic variant called MI210. The third generation expanded this into a multi-chip system with accelerator complexes and input-output dies. A future fourth generation is planned for 2025, featuring new models in both single-card and multi-GPU platform forms designed for universal baseboards.
Why it matters
CDNA marked a strategic split in AMD graphics design, separating consumer gaming technology from professional compute needs. This allowed the company to optimise hardware specifically for data centre efficiency and throughput. The shift to multi-chip modules enabled significant scaling of transistor counts and memory bandwidth without the yield issues associated with extremely large single dies. It established a foundation for competitive acceleration in the enterprise market.
Questions people ask about CDNA
What is the main difference between CDNA and RDNA?
CDNA is designed for data centre compute tasks and removes graphics-specific hardware like display engines. RDNA focuses on consumer gaming graphics. Both share some underlying technology but target different markets and workloads.
Which manufacturing process is used for CDNA chips?
TSMC manufactures CDNA chips. The first generation used the N7 node, the second used N6, and the third generation combines five-nanometre and six-nanometre process technologies for its various chiplets.
How does CDNA handle memory bandwidth?
It uses High Bandwidth Memory (HBM) stacks connected via a wide memory bus. Bandwidth has increased across generations, from roughly 1200 GB/s in the first version to higher speeds in later iterations using HBM2e and HBM3 technologies.
What products use the CDNA architecture?
The AMD Instinct series of accelerator cards uses CDNA. Specific models include the MI100, MI250X, MI250, MI210, and the MI300 series. These are used in servers and high-performance computing systems.
Good reading and sources
- Official websiteOfficial site
- Checking your browserAnandTech, Article
- Hot Chips 34Chips and Cheese, Article
The facts on this page come from the Wikipedia article "CDNA (microarchitecture)", 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.
Like CDNA
All graphics processorsMore from AMD
All of them
🇺🇸 United StatesApple M1
ARM-based system-on-a-chip for Macs and iPads, designed by Apple and manufactured by TSMC.SoC · Arm · 2020
🇺🇸 United StatesApple A14 Bionic
Apple's 5 nm phone and tablet chip with six CPU cores and a 16-core Neural EngineSoC · Arm · 2020