For the first time, the NVIDIA Ampere ampere family used two manufacturing processes, the data center-oriented, deep-learning A100 is TSMC 7nm and the game-oriented RTX 30 series is Samsung 8nm. From the current situation, Samsung 8nm performance is indeed general, regardless of the yield or performance are poor, the most direct manifestation is that the RTX 30 series has always been seriously inadequate supply, and frequency, almost impossible to overclock.
On the other hand, the AMD RX 6000 series, which uses TSMC’s 7nm process, greatly improves frequency, performance and energy efficiency while the process remains unchanged.
Whatever reason NVIDIA chose Samsung 8nm, worried about TSMC’s lack of capacity, or Samsung’s lower offer, this generation is not a success.
As for the Ampere game cards that some say will be a generation of 7nm processes, it’s obviously very unlikely, which is equivalent to redesigning the streaming film once, cost and time are not allowed.
Looking ahead, NVIDIA’s next-generation GPU, codenah “Hopper,” is taken from Grace Hopper, the first lady of computer software engineering and mother of Cobol, the compilation language.
According to the information obtained by the exposure master kopite7kimi, the Hopper generation will all be on TSMC 5nm, including the data center-class H100 core, the game-class GH20x series core.
But he also left a bottom, saying it was too early and that everything could change.
Curiously, the core code of the NVIDIA game card is the 10x series after the new architecture, and I don’t know why it’s just the 20x series.
In fact, in March this year kopite7kimi once said that Hopper would TSMC 5nm, just like apple A14, when the code name used was GH100.
Has Hopper quietly changed to the second edition?