AMD Instinct MI100
CDNA architecture GPU for high-performance computing and AI workloads.
The AMD Instinct MI100 is a data centre accelerator based on the new CDNA architecture. Launched in late 2020, it targets supercomputers and AI tasks with 32GB of memory. It uses a 7nm process and connects via PCIe Gen4 or Infinity Fabric links.
- First GPU built on the dedicated CDNA architecture
- Features 128 compute units and 32GB HBM2 memory
- Manufactured on TSMC 7nm process node
- Supports Infinity Fabric for fast multi-GPU linking
- Targets exascale supercomputing and AI workloads
What it is
The AMD Instinct MI100 is a graphics processing unit designed for high-performance computing and artificial intelligence. It marks the debut of the CDNA architecture, which separates data centre needs from gaming graphics. The chip offers 32GB of high-bandwidth memory and fits into standard server slots. It aims to provide strong floating-point performance for scientific simulations and machine learning tasks.
How it came about
AMD launched this accelerator at the SC20 supercomputing conference in November 2020. It followed the Instinct MI50 and MI60 models based on older graphics architectures. The company split its GPU development into RDNA for gamers and CDNA for servers. This move allowed engineers to remove game-rendering logic and focus on raw compute density. The design supports major exascale supercomputer contracts won by AMD.
Inside the chip
The chip uses the Arcturus die with 128 compute units. These units are optimized for floating-point and integer operations rather than graphics rendering. It connects to host processors using PCIe Gen4 interfaces. For multi-GPU setups, it employs Infinity Fabric links to boost communication speed between cards. The architecture supports four GPUs per CPU socket in hive configurations. Software support comes through the ROCm platform, which helps port applications from other ecosystems.
How it was made
TSMC manufactures the Arcturus GPU using a 7-nanometre process node. The card draws up to 300 watts of power. It carries 32GB of HBM2 memory, which AMD selected as a cost-effective balance for high-performance computing users. The physical form factor is a standard PCIe card, allowing integration into existing server infrastructure without custom backplanes.
Where it went
System builders like HPE, Dell, Supermicro, and Gigabyte integrate these accelerators into their servers. They pair with AMD EPYC 7003 series processors. The hardware powers large-scale scientific simulations and AI training workloads. Early adopters include national laboratories working on exascale systems such as El Capitan. The card serves as a primary compute engine in data centres requiring high throughput.
Versions and family
This model stands as the single SKU for the initial CDNA launch. It replaces the previous Instinct MI50 and MI60 cards. Unlike earlier generations that had varying memory sizes, this version standardizes on 32GB. It competes directly with NVIDIA A100 accelerators in the high-end market. Future iterations are expected to build upon this architectural foundation.
Why it matters
The release helped AMD secure roles in leading exascale supercomputing projects. It demonstrated that domain-specific architectures could challenge general-purpose GPU designs. The shift to CDNA signaled a long-term commitment to the data centre market. By offering competitive performance per dollar, it attracted attention from cost-conscious enterprise buyers. The accompanying software updates aimed to reduce barriers for developers switching platforms.
Problems and reception
Reviewers noted the strong theoretical performance gains over previous AMD generations. Comparisons focused on its value proposition against rival chips with higher memory capacities. Some observers pointed out that software ecosystem maturity remained a key challenge. The decision to limit memory to 32GB was seen as a strategic cost control measure. Industry analysts viewed it as a significant step in closing the gap with market leaders.
Questions people ask about AMD Instinct MI100
What architecture does the MI100 use?
It uses the CDNA architecture, which AMD designed specifically for data centre and high-performance computing tasks. This separates it from the RDNA architecture used in gaming graphics cards.
How much memory does the chip have?
The Instinct MI100 includes 32GB of HBM2 memory. AMD chose this capacity to balance performance and cost for high-performance computing customers who prioritize bandwidth.
Who manufactures the MI100 GPU?
Taiwan Semiconductor Manufacturing Corp (TSMC) produces the Arcturus GPU die using its 7-nanometre process technology. AMD designs the chip and handles the final assembly.
Can multiple MI100 cards work together?
Yes, the cards support Infinity Fabric Link. This allows up to four GPUs to connect in a hive configuration for faster communication than standard PCIe connections provide.
Good reading and sources
There is no English Wikipedia article on AMD Instinct MI100 yet, so the facts on this page come from the sources listed above: servethehome.com/amd-radeon-instinct-mi100-32gb-cdna-gpu-launched/, nextplatform.com/compute/2020/11/16/amd-at-a-tipping-point-with-instin. 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.
Like AMD Instinct MI100
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