Nvidia H100
AI and HPC accelerator based on the Hopper architecture
The Nvidia H100 is a graphics processing unit designed for data centres, built on the Hopper architecture. It accelerates artificial intelligence training and high-performance computing workloads using fourth-generation Tensor Cores and a dedicated Transformer Engine.
- Uses Hopper architecture with dedicated Transformer Engine
- Delivers 60 teraflops of FP64 computing performance
- Supports FP8 precision for faster AI inference
- Features hardware-based confidential computing security
- H100 NVL variant includes 188GB HBM3 memory
What it is
The Nvidia H100 is a graphics processing unit intended for enterprise data centres. It uses the Hopper microarchitecture to handle large-scale artificial intelligence and high-performance computing tasks. The chip includes a dedicated Transformer Engine to support trillion-parameter language models. It offers fourth-generation Tensor Cores that support various precisions, including FP8, to improve speed and reduce memory usage during inference and training.
Inside the chip
The H100 features fourth-generation Tensor Cores and a Transformer Engine with FP8 precision. It delivers 60 teraflops of double-precision computing power. New DPX instructions accelerate dynamic programming algorithms like DNA sequence alignment. The chip supports Multi-Instance GPU technology, allowing managers to partition resources securely. It also includes hardware-based confidential computing to protect data integrity within a trusted execution environment.
Where it went
This processor powers systems for conversational AI, large language model training, and scientific research. It accelerates genome sequencing and protein structure prediction. The H100 NVL variant supports models with up to 70 billion parameters using 188GB of HBM3 memory. It connects via fourth-generation NVLink and NDR Quantum-2 InfiniBand networking to scale across massive clusters.
Versions and family
The H100 NVL is a PCIe-based variant with an NVLink bridge and 188GB of memory. It pairs with the Grace CPU in the Grace Hopper superchip, using a fast chip-to-chip interconnect. The standard H100 serves as the predecessor to the Nvidia Blackwell architecture.
Why it matters
The company states the H100 speeds up large language model training by 30 times compared to prior generations. It claims to offer four times faster AI training for GPT-3 models. The chip enables exascale high-performance computing and brings trillion-parameter AI within reach of researchers.
Questions people ask about Nvidia H100
What architecture does the Nvidia H100 use?
The H100 is built on the Nvidia Hopper architecture. This design includes breakthrough innovations such as a dedicated Transformer Engine and fourth-generation Tensor Cores to accelerate artificial intelligence workloads.
How much memory does the H100 NVL have?
The PCIe-based Nvidia H100 NVL variant utilizes 188GB of HBM3 memory. This configuration helps provide optimum performance for large language models with up to 70 billion parameters.
What is the predecessor to the H100?
The Nvidia A100 is the predecessor to the H100. The H100 offers significant performance improvements over the A100, including faster training speeds and new DPX instructions for specific algorithms.
Good reading and sources
There is no English Wikipedia article on Nvidia H100 yet, so the facts on this page come from the sources listed above: nvidia.com/en-us/data-center/h100/, tomshardware.com/news/nvidia-h100-hopper-benchmark-results-published. 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.