At the upcoming GTC 2022, NVIDIA CEO Jen-Hsun Wong will have a keynote speech (March 22 at 8 a.m. PST), which is likely to bring a new generation of Hopper architecture GPUs. In the past two days, a lot of details about the Hopper architecture GPU have been leaked.
According to VideoCardz, the leaked images show that the GH100 supports up to 6 high-bandwidth memory stacks and is still a monolithic structure, although its specific specifications are still uncertain. Perhaps, as rumored, the GH202 will be available in a multi-chip modular design (MCM) and CoWoS advanced package.
NVIDIA will introduce a number of products based on the GH100, including the SXM form factor H100 compute card for DGX motherboards, the DGX H100 workstation, the DGX H100 SuperPod, and the H100 compute card with PCIe interface.
Previously, some netizens sorted out the information stolen by the previous hackers and circulated information, and summarized the situation of GH100 based on the Hopper architecture. The GH100 is said to be manufactured using the TSMC 5nm process and will have a 48MB L2 cache, which is an improvement over the Ampere architecture GA100's 40MB, which is three times (16MB) for the AMD Instinct MI250. However, compared to the 96MB of the Ada (Lovelace) architecture's flagship chip AD102, the GH100's L2 cache is half less.
It is rumored that GH100 is equipped with 8 sets of GPCs, each GPC is equipped with 9 sets of TPC, and there are two sets of SM in each TPC, if the number of CUDA cores per group of SM does not change, it means that there are a total of 144 sets of SM and 18432 CUDA cores. Only 1 of the 8 GPC groups has a 3D engine, and the other 7 groups will not be equipped. Although the GH100 is configured with 144 sets of SM, it should not be fully enabled, and Nvidia is expected to disable 15% to 20% of SM.