Hopper servers typically refer to servers equipped with NVIDIA Hopper architecture GPUs. The NVIDIA Hopper architecture is specifically designed for accelerated computing, supporting significant improvements in AI, high-performance computing (HPC), memory bandwidth, interconnects, and communication, achieving nearly 5TB/s of external interconnect bandwidth. Hopper architecture GPUs, such as the H100, are the first to support PCIe 5.0 and utilize HBM3, enabling 3TB/s of memory bandwidth.

The NVIDIA HGX H100 is a key GPU server building block powered by the NVIDIA Hopper architecture, offering low-latency high performance and integrating a full suite of features from networking to data center-level computing. The HGX H100 8-GPU is a critical component of the next-generation Hopper GPU servers, featuring eight H100 Tensor Core GPUs and four third-generation NV switches. Each H100 GPU is connected to all four NVLink switches via multiple fourth-generation NVLink ports for high-speed communication.

Additionally, the Hopper architecture includes features such as multi-chip module (MCM) design, evolved Tensor Cores, and software ecosystem support, enabling Hopper servers to deliver higher application performance in AI training, inference, data analysis, and HPC applications.

$Broadcom(AVGO.US) $NVIDIA(NVDA.US)  $Amazon(AMZN.US)