Nvidia released the H100 with the new Hopper architecture in late March, which has the strongest GPU specifications currently available at NVIDIA. The core architecture of NVIDIA H100 is similar to that of the previous generation Ampere, with mathematical operations arranged on 144 CUDA sets, with up to 18432 FP32 (single precision), 9216 FP64 (double precision) CUDA cores, supplemented by 576 fourth-generation Tensor cores.
NVIDIA's white paper, which was unveiled in early May, gave the outside world a better understanding of the Hopper architecture. The H100 core is manufactured by TSMC's N4 process, with 80 billion transistors built in and a core area of only 814m㎡. For comparison, the previous generation of A100 cores are manufactured using TSMC's 7nm process, with 54.2 billion transistors built into the 826m㎡ core.
As a GPU for professional computing, the H100 uses HBM3 high-bandwidth memory, and NVIDIA has six HBM3 high-bandwidth memory stacks on either side of the core. The core has a built-in 5120-bit HBM3 memory bit width, NVIDIA can configure up to 80GB of video memory, SXM5 version (HBM3 memory) bandwidth is 3TB/s, PCIe version (HBM2e) is 2TB/s.
The H100 host interface has also ushered in an upgrade, and the SXM form factor PCB board is equipped with a new generation of NVLink, with a bandwidth of 900GB/s. The AIC plug-in version uses PCIe 5.0 x16 (with 128GB/s) interface, both of which introduce resource-pooling to speed up data exchange between GPUs.
Although the H100 has 144 groups of cells, the SXM version also only enables 132 of them. The PCIe version is only 114 groups, and the highest frequency of both is 1.8GHz. Not only that, the power of the H100 core is as high as 700W, the PCIe version has reached 350W, and the previous generation A100 is only 400W; While improving performance, the H100's power consumption is also rising dramatically.
Editor's Comments: The white paper of NVIDIA H100 analyzes the core architecture of H100 in more detail, and analyzes the internal architecture and design concept of the core. The announcement of the H00 core has a high reference value for the Ada Lovelace that will arrive in the second half of the year, and NVIDIA has a high probability of fine-tuning the architecture on the basis of the CUDA of the H100 to adapt to the needs of the game.