NVIDIA’s new Hopper GPU architecture to be announced at GTC 2022

NVIDIA is giving some hints of what will be presented in terms of new GPUs and Hopper architecture at GTC 2022. Nvidia CEO Jensen Huang has a habit of being quite mysterious about new technologies Nvidia is planning. and he likes to play a little with the press. As far as we know, NVIDIA’s Hopper GPUs were supposed to be the company’s first MCM (Multi Module System) based GPUs and could have up to 2 GPUs in a single package.

Arguably, an MCM-based design is the next step in the evolution of GPUs, considering that we are now limited by the size of microarchitectures. Architectural and design enhancements to MCM are the next stepAnd since AMD is already working on all of this with Ryzen processors, it makes sense that GPUs are the next step in this big plan, which would explain why NVIDIA wants to be one step ahead with this technology.

The speculative specifications of the upcoming NVIDIA H100 GPU (Top of the Range in the new generation) is a 2 x 900 mm² die based on the TSMC N5 process. This does a total of 140 billion transistors or 70 billion per GPU. This leads to an increase in density of 18% per square mm. It will feature 268 GPU clusters and a total of 34,304 CUDA cores.

The memory bus will be 6144 bits and a total of 128 GB of HBM3 can be integrated. The TDP will certainly be huge and in consumption by 500W+. Unfortunately, we don’t know much else at the moment and a leaker recently claimed that the Hopper architecture would remain monolithic (at least as far as the H100 chip is concerned) and not be based on MCM. As the information is not official, it remains to wait for the announcement from Nvidia.

NVIDIA architectures are always based on pioneering computer scientists. Nvidia’s Hopper architecture is based on Grace Hopper, who was one of the pioneers of computing and one of the first programmers for the Harvard Mark 1 project. She also popularized the idea of ​​programming languages ​​independent of the development of COBOL, one of the first high-level programming languages ​​still in use today.