NVIDIA Unveils Hopper Architecture and H100 GPU for Next-Gen AI
NVIDIA introduces the Hopper architecture and its first H100 GPU, delivering a massive performance leap for AI data centers. The new chip features a specialized Transformer Engine to accelerate large language models.
NVIDIA reveals its new Hopper architecture at GTC to power the next generation of AI data centers. Named after pioneering computer scientist Grace Hopper, this new architecture succeeds the previous Ampere architecture and delivers an order of magnitude performance leap for accelerated computing workloads.
The first Hopper-based GPU, the NVIDIA H100, contains 80 billion transistors and stands as the world's largest and most powerful accelerator. It introduces a specialized Transformer Engine that speeds up natural language processing networks by up to six times without losing accuracy, making it ideal for giant AI language models, genomics, and complex digital twins.
Built on a cutting-edge TSMC 4N process, the H100 is the first GPU to support PCIe Gen5 and utilize HBM3 memory for incredible 3TB/s bandwidth. The chip also features second-generation Secure Multi-Instance GPU technology that allows a single GPU to be securely partitioned into seven isolated instances for efficient cloud computing.