Source: IT House
IT House news on March 22, Videocardz once again leaked the entire contents of the upcoming Hopper H100 GPU on the eve of NVIDIA GTC. IT House reminds that this is a new generation of GPUs for data centers and NVIDIA's first 5nm product.
From the exposure content, we can see that NVIDIA's new-generation Hopper H100 GPU is still a huge monolithic structure, so it breaks the speculation on the Internet that the new generation of graphics cards will adopt MCM design, and 144 separate streaming multiprocessors (SM) can be calculated from this.
▲ The picture shows the NVIDIA Hopper GPU, a single chip, containing VRM and HBM
From the figure, the chip supports up to 6 high-bandwidth HBM memory stacks, but since the full specifications have not yet been revealed, we can't go any further.
NVIDIA will also launch its new data center/server-based products, which foreign media believe:
NVIDIA will introduce several GH100 GPU-based products, such as the SXM-based H100 card for DGX motherboards, the DGX H100 workstation, and even the DGX H100 SuperPod. In addition, Nvidia will also release PCIe-based H100 products.
Foreign media assumes that if its SM to CUDA core ratio is the same as Turing card, then each SM has exactly 64 CUDA cores, a total of 9126 CUDA cores; assuming that its base frequency is 2.2 GHz, which can provide at least 40 TFLOP double precision performance.
In addition, the number of VRMs on the PCB also indicates that the power consumption of this card will be very large, considering the revelations about Ada Lovelace. At present, this thing is definitely a performance beast, and so is power consumption, and we expect to see NVIDIA CEO Jen-Hsun Huang officially bring it to everyone tomorrow.
NVIDIA's GPU architecture has always been named after a pioneer in computing, and this architecture (Hopper) is based on Grace Hopper, one of the first programmers at Harvard Mark 1, the inventor of the first linker, and who also promoted the development of the COBOL language.