Recently, the NVIDIA GTC conference was held with great momentum, and CEO Huang Renxun gave a keynote speech, introducing NVIDIA's latest research and development progress in AI computing hardware, software ecology, and edge intelligence. Rocket! At this meeting, Nvidia announced the launch of the new generation GPUBlackwell and later this year, the first supercomputing chip based on the Blackwell architecture - GB200.
It is worth noting that a GB200 accelerator card combines two B200GPUs and an independent GraceCPU, which can improve the inference efficiency of LLM (Large Language Model) by 30 times. Compared to H100, it can "reduce cost and energy consumption by 1/25".
The launch of larger computing power GPUs and the emphasis on power consumption, on the one hand, demonstrate a continuous surge in demand for AI computing power, and the computing infrastructure industry chain is expected to continue to benefit. On the other hand, technological changes will also give rise to three new opportunities in the industry chain.
Optical module: It is understood that the GB200 will adopt a 1.6T optical module. In the future, if the GB200 can continue to scale up and improve its computing speed, 1.6T or even higher optical modules are expected to become a major trend. With the accelerated penetration of new technologies, the industry may form a trend of leading concentration.
High speed copper cable connection: Nvidia provides GB200NVL72 servers for enterprises, which use high-speed copper cable interconnection internally, with a cumulative cable length of nearly 2 miles and a total of 5000 independent copper cables. Compared to optical modules, high-speed copper cables are more suitable for short distance and cost sensitive application scenarios. Domestic companies have already established relevant technological layouts, and electronic connector representative enterprise, Luxon Precision, stated that copper connections have always been the core product of Luxon's communication business. Currently, the company has developed multiple technologies based on its independently developed Optamax bulk cable technology, achieving breakthroughs in copper physical characteristics.
Huaxia Fund believes that in the long term, it can focus on 5G communication ETFs (515050), with nearly 40% equity in optical modules, optical communication, and AI computing power concepts. The top ten holding stocks are Lixun Precision, ZTE Communications, Zhongji Xuchuang, Industrial Fulian, Zhaoyi Innovation, Sanan Optoelectronics, Goer Group, Ziguang Group, Zhuoshengwei, and AVIC Optoelectronics.
Liquid cooling: With the significant improvement of AI chip performance, power consumption is also expected to rise. According to AMD, the current air cooling can meet the cooling needs of GPUs with power of 600-700W, and the cooling problem will become more prominent when operating above 1000W. It is understood that GB200NVL72 servers will use an integrated water-cooled cooling solution, all of which will adopt liquid cooled MGX packaging technology, reducing costs and energy consumption by 25 times. Liquid cooling is expected to become a new trend in cooling.