NVIDIA's Monstrous GH100 Hopper GPU Might Pack More Than 140 Billion Transistors
In case you haven't been following high-performance computing (HPC) news, NVIDIA is sort of ruling the roost lately. Its massively-parallel processing engines like GV100 and GA100 are the fastest chips around for computing many types of crunchy math operations, big data analytics, AI etc. They achieve that supreme speed in-part by simply being absolutely massive, packing a gigantic number of processor cores into a single die.
Rumor has it that the company's next HPC-focused product, code-named "Hopper" (after one of the matriarchs of computing, American mathematician and computer scientist Grace Hopper), will be even larger than those chips in its top-tier form. In fact, it's expected to be so big that many people initially assumed it would be a multi-chip module (MCM) design. The latest leaks specifically state that that's not the case, but of course, one must take all of these rumors with a grain of salt.
Original Chinese above, and Google Translated below.
So it goes with the newest little nugget of information about GH100, speculated to be the largest Hopper chip. A poster named Zhangzhonghao over on the Chinese-language Chiphell forums, a frequent source of just this sort of leak, has stated rather vaguely that NVIDIA's GH100 GPU could comprise some 140 billion transistors. If we look at this alongside the aforementioned previous leaks stating that this is still a monolithic chip, it will be nearly three times as dense as NVIDIA's Ampere-based GA100.
Comparing to the competition, Intel's upcoming Ponte Vecchio Xe-HPC accelerator is said to have over 100 billion transistors, while AMD's Instinct MI200 uses a pair of dice with some 58 billion transistors each. The transistor count seems sky-high not only in comparison to the competition and NVIDIA's own last-gen, but also because of the expected specifications of the chip, which is supposed to have around 25% more shader modules than the GA100. Perhaps Hopper has some unknown new functionality mandating all those extra transistors.
