According to the latest report from insider @kopite7kimi, the upcoming NVIDIA HPC accelerator called GH100 will have a massive die size of just under 1000 mm2. This means the GPU can be up to 20% larger than the GA100 based on the Ampere architecture.
The other day there was information about the dispute over the trademark “Hopper”, which NVIDIA is facing from a company using a similar name for its products. In the event that NVIDIA’s ongoing negotiations with a company called “Dish Network” do not conclude in the next few weeks, NVIDIA may not be able to announce its new GH100 GPU at GTC 2022 in March.
Also, it looks like the rumors about the GH100 multi-chip module design were wrong. Now, two leakers, including @kopite7kimi and @greymon55, are claiming that the GH100 is actually a one-piece die design. However, it appears that this is indeed an MCM bunker design. This chip is supposedly called GPU GH102.
The GH100 is also expected to be a very power hungry chip. Whistleblower @kopite7kimi previously reported that TDP could even reach 1000W, which would require a very efficient cooling system. However, the company already offers a series of water-cooled datacenter GPUs called the HGX, so technically this shouldn’t be a problem.