NVIDIA’s Hopper architecture GH100 GPU has 140 billion transistors
The Hopper architecture is NVIDIA’s first multi-chip module design (MCM)-based GPU, which will be manufactured using TSMC’s 5nm process and CoWoS advanced packaging, support for HBM2e and other connectivity features, aimed at the data center, is expected to debut in mid-2022, competing against Intel’s Xe-HP architecture and AMD’s CDNA 2 architecture.
According to Wccftech, the number of transistors of the GH100 based on the Hopper architecture will reach 140 billion, that’s almost 2.5 times the current Ampere-based GA100 (54.2 billion) or AMD’s CDNA 2-based Instinct MI200 series (58 billion). The chip size of the GH100 is said to be close to 900mm², which is smaller than the previously rumored 1000mm², but larger than the GA100 (862mm²) and the Instinct MI200 series (about 790mm²). It is rumored that the GH100 is equipped with a total of 288 SMs, which can provide three times the performance of the current A100 computing card.
It is rumored that Nvidia also has a secret weapon, which is the COPA solution based on the Hopper architecture. Nvidia researchers have previously published an article detailing how Nvidia is exploring how to deploy multi-chip designs for future products, which mentioned related designs. Nvidia will have two designs based on the same architecture, but for the high-performance computing (HPC) and deep learning (DL) segments. HPC will use the standard scheme, DL will have a huge independent cache connected to the GPU.
Nvidia will likely reveal the Hopper architecture GPU at GTC 2022, where it will be available in more detail.
