What is the NVIDIA Hopper Architecture?
2026-10-28
The NVIDIA Hopper architecture entered the market in 2022 and, as of 2026, remains one of the most widely used GPU microarchitectures in modern data centers focused on AI and machine learning. It was developed in response to emerging computational demands, with large neural networks, large language models (LLMs), and generative AI in mind from the outset. Hopper is also widely used for demanding analytics and high-performance computing (HPC).
The best-known and most widely deployed model in the family is the NVIDIA H100, based on the GH100 GPU. It is available in PCIe and SXM form factors. Hopper-based accelerators feature fourth-generation Tensor Cores and support FP64, FP32, TF32, BF16, FP16, FP8, and INT8 precisions. The H100 also introduced NVIDIA’s Transformer Engine, which uses FP8 to accelerate the training and inference of transformer-based models.
The flagship NVIDIA H100 uses high-bandwidth HBM memory. Depending on the model, memory bandwidth reaches up to 3.35 TB/s per GPU, while an H100 NVL configuration with two accelerators can provide aggregate bandwidth of up to 7.8 TB/s. Hopper also supports fourth-generation NVIDIA NVLink, providing up to 900 GB/s of bidirectional GPU-to-GPU bandwidth.