Nvidia Unveils Hopper H100 Datacenter GPU With 80 Billion Transistors
Nvidia introduces the Hopper H100 GPU for datacenters, packing 80 billion transistors on a custom TSMC 4N process to outperform the previous A100. The new chip focuses heavily on AI and supercomputing workloads with massive upgrades to memory and interconnect bandwidth.
Nvidia reveals its new Hopper architecture and the H100 GPU at the GPU Technology Conference, targeting datacenters and supercomputers rather than consumer graphics. The H100 succeeds the Ampere A100 and faces increasing competition from AMD's Instinct lineup, pushing Nvidia to significantly boost AI and high-performance computing capabilities. This new processor clearly separates itself from the upcoming Ada architecture that will eventually power consumer GeForce cards.
The H100 GPU features a massive 80 billion transistors manufactured on a custom TSMC 4N process, representing a substantial increase over the A100's 54 billion transistors. Nvidia equips the chip with 80GB of HBM3 memory that delivers 3 TB/s of bandwidth, providing a 50% speed boost over the previous generation's memory. While the company keeps core counts and clock speeds under wraps, the sheer density of the chip highlights a major leap in hardware capability.
Connectivity sees major upgrades with the inclusion of fourth-generation NVLink technology that reaches up to 900 GB/s of bandwidth, offering 1.5 times the throughput of the older A100. For systems that do not utilize NVLink, the H100 supports PCIe 5.0, which doubles the maximum bandwidth to 128 GB/s compared to PCIe 4.0. These combined improvements in memory speed and data transfer position the H100 as a powerful new tool for massive AI workloads.