The Hopper graphics architecture will most likely be announced next week during the speech of the NVIDIA Director General. According to rumors, this is a long-awaited series of products intended for data centers and high-performance computing.
NVIDIA Hopper is considered the first multiple module of the company, which includes as many as two chiploads. In addition, it is rumored that it is one of the largest graphics processors ever created by NVIDIA, with the size of one crystal exceeding the grid limit of 853 mm². If other rumors are true, the GH100 graphics processor can also have up to 140 billion transistors, which is almost 2.6 times more than its predecessor ampere.
The complete configuration of the GPU chipboard can offer up to 144 streaming multiprocessors and 18,432 kernels FP32 (CUDA). This means that the design with two chiplot will have up to 288 SM and 36,864 kernels. Naturally, the real product (such as H100 Tensor Core) will undoubtedly have trimmed characteristics to match the required performance and performance. At its predecessor, 16% of the chip was turned off and there was no one memory module.
In addition, the GH100 is not the only GPU Hopper, which is now working NVIDIA. According to the leaked data published by the hacker group, which was able to access NVIDIA servers, the GH202 graphics processor is also developed. However, at the moment, nothing is known about this particular chip, except that it will be released later.
The speech of the NVIDIA Director General will broadcast on March 22 at 8:00 on Pacific.
Source: NVIDIA / VideoOcardz-Com