A100
Data centre GPU for AI and high-performance computing.
The A100 is a graphics processing unit designed by Nvidia for data centres. Launched in May 2020, it uses the Ampere architecture to accelerate artificial intelligence, analytics and scientific simulations. An 80 GB version followed in November 2020 with faster memory bandwidth.
- Launched in May 2020 with Ampere architecture
- Supports partitioning into seven GPU instances
- 80 GB model offers over 2 TB/s memory bandwidth
- Consumes between 400 W and 500 W of power
What it is
The A100 is a graphics processing unit built for enterprise data centres. Nvidia introduced it in May 2020 as part of its Ampere architecture family. The chip handles heavy workloads such as artificial intelligence training, data analytics and high-performance computing simulations. It supports partitioning into seven independent instances to share resources among multiple users. An updated model with 80 GB of memory arrived in November 2020, offering higher bandwidth for large datasets.
How it came about
Nvidia announced the A100 based on the GA100 GPU in May 2020. The company later released an enhanced version with 80 GB of HBM2e memory in November 2020. This expansion aimed to support larger models and faster data throughput. The A100 serves as the core of the Nvidia data centre platform, preceding the Hopper architecture which was expected to follow in 2022.
Inside the chip
The chip uses the Ampere microarchitecture and includes Tensor Cores for matrix calculations. It supports various precision formats from FP32 down to INT4. A key feature is Multi-Instance GPU technology, which splits one physical chip into up to seven virtual GPUs. This allows different tasks to run simultaneously with guaranteed performance levels. The design also supports structural sparsity to improve inference speed without changing code.
How it was made
The A100 is manufactured using processes associated with TSMC, though specific node details are not listed in the provided text. The 80 GB variant uses HBM2e memory modules. Power consumption ranges from 400 watts for standard air-cooled setups to 500 watts for custom liquid-cooled solutions. The chips connect via PCIe slots or SXM4 server boards, often linked with NVLink bridges for multi-GPU systems.
Where it went
This processor powers elastic data centres for AI and scientific research. It accelerates training for complex models like BERT and deep learning recommendation systems. In high-performance computing, it reduces simulation times for applications such as Quantum Espresso and GROMACS. Data scientists use it for big data analytics with tools like RAPIDS and Apache Spark. It also handles inference tasks for conversational AI and speech recognition.
Versions and family
Two main memory configurations exist: a 40 GB model and an 80 GB model. The 80 GB version offers double the memory capacity and higher bandwidth. Other Ampere-based cards include the A40, A6000, A10, A30 and A2, but these use different GPU dies like GA104 or GA102. The A100 specifically relies on the GA100 die. It succeeds the Volta generation and is followed by the Hopper series.
Why it matters
The A100 set performance records in MLPerf benchmarks for AI training and inference. It enabled a world record where 2,048 GPUs solved a BERT workload in under a minute. The chip provides significant speedups over previous generations, with claims of up to 20 times higher performance than Volta. Its ability to partition resources helps maximise infrastructure utilisation in cloud environments.
Questions people ask about A100
What is the Multi-Instance GPU feature?
MIG allows a single A100 chip to be split into up to seven independent GPU instances. This lets multiple users access acceleration simultaneously while ensuring each job gets guaranteed quality of service.
How much memory does the A100 have?
The A100 comes in two versions. The standard model has 40 GB of memory, while the enhanced version launched in November 2020 features 80 GB of HBM2e memory for larger datasets.
What types of workloads is the A100 designed for?
It is built for artificial intelligence training and inference, high-performance computing simulations, and large-scale data analytics. It handles tasks like natural language processing and scientific modelling.
Good reading and sources
There is no English Wikipedia article on A100 yet, so the facts on this page come from the sources listed above: nvidia.com/en-us/data-center/a100/, nextplatform.com/ai/2021/12/06/stacking-up-amd-mi200-versus-nvidia-a10. 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.