Computex 2023 was hosted at Taipei Nangang Exhibition Center from May 30 through June 2 and DIGITIMES Research paid a visit to the show to survey and outline the trends that new AI servers showcased at the event, including the increasing popularity of models powered by Nvidia's HGX H100 among the suppliers; AI servers with PCIe support trending toward high-density and high-application-flexibility development; servers powered by Nvidia's GH200 boasting the best performance running Nvidia's Large Language Model (LLM) and Nvidia's release of MGX reference server designs.
With enterprises' dependency on cloud computing continuing to grow during the post-pandemic era and the ongoing tech competition between the US and China, semiconductors have grown to become the most important sector of the IT industry, while ChatGPT's emergence also spurred a frenzy for generative AIs. With the two becoming the hottest topics within the IT industry, AI servers surely also became one of the main spotlights of the show.
OEMs/ODMs' rack servers powered by the HGX H100 took center stage at Computex 2023. Since rack servers equipped with eight H100 GPUs that feature Nvidia's latest SXM5 socket are considered the best choice to run LLM training and inferencing, ODMs such as Supermicro, Giga Computing, and Quanta Cloud Technology (QCT) have all launched related server products and are expected to see tight supply due to their robust demand recently, DIGITIMES Research's latest report shows.
AI servers with PCIe support are lower in cost but have high configuration flexibility and compatibility with conventional server racks, so they remain one of the favorite options enterprises pick for their deployment of AI applications.
DIGITIMES Research's observation of the AI servers with PCIe support exhibited by leading manufacturers showed that the manufacturers' AI server development is mainly focusing on featuring high computing density and a variety of chassis designs to accommodate different application scenarios.
Nvidia CEO Jensen Huang, during his keynote at the show, noted that Grace Hopper Superchip unveiled in March 2022 had already begun volume production and demonstrated the Nvidia DGX GH200 AI supercomputer that is powered by Grace Hopper Superchips. Nvidia also announced MGX reference server designs, aimed at creating system designs to accelerate time-to-market for AI servers and strengthening its influence on OEMs and ODMs.
