NVIDIA has officially launched the H100 GPU, the company’s most powerful AI training chip to date. Built on the latest Hopper architecture, the H100 introduces 4th Generation Tensor Cores, specifically designed to accelerate training and inference for large-scale artificial intelligence models, including generative AI and deep learning systems.
The H100 GPU supports the FP8 precision format, a key innovation that significantly reduces computation time and power consumption while preserving model accuracy. NVIDIA reports that the H100 can outperform its predecessor (the A100) by up to 40% in training performance for models such as GPT-style transformers, computer vision networks, and reinforcement learning systems.
With a higher memory bandwidth (3 TB/s) and a massive 80 GB of HBM3 memory, the H100 is engineered for data centers, high-performance computing (HPC) environments, and cloud platforms. It is already being adopted by major AI companies and cloud providers, including AWS, Microsoft Azure, and Google Cloud.
One of the most exciting updates is the NVLink 4.0 interconnect, allowing multiple H100 GPUs to communicate with ultra-low latency. This makes it easier to scale AI models across clusters of GPUs, critical for organizations training models with hundreds of billions of parameters.
NVIDIA CEO Jensen Huang emphasized during the keynote:
“The H100 is the engine behind the AI factories of the future. It enables breakthroughs in language understanding, robotics, scientific research, and every industry touched by AI.”
The H100 is available as part of the new NVIDIA DGX H100 server and will also power NVIDIA’s AI supercomputer clusters. Early access to the hardware began in Q2 2025, with wider availability expected throughout the second half of the year.
With the rapid rise of generative AI, the demand for high-performance AI chips is higher than ever. The H100 positions NVIDIA to remain at the forefront of the AI hardware industry.
