Under the wave of generative artificial intelligence, how does the ICT industry play AI?

On July 6, the 2023 World Artificial Intelligence Conference (WAIC) opened in Shanghai, and AI large models are one of the hot topics. Baidu, Ali, Huawei, Tencent, SenseTime, JD.com, Netease, Fourth Paradigm and more than ten large models are showing AI muscles on the spot.

Hou Yang, Senior Vice President of Microsoft and Chairman and CEO of Microsoft Greater China, said in his speech that every company needs to become a digital company in the future, and every application of the company will be driven by artificial intelligence.

Under the sweep of AIGC, ICT vendors are also accelerating their entry, and industry chain leaders such as Huawei, New H3C, and FII are iterating on software and hardware.

New H3C recently released the private domain large model "Baiye Lingxi" LinSeer, an AI server specially designed for large model training, an 800G CPO silicon photonics data center switch, and upgraded the Aofei computing power platform.

You Xuejun, co-president, chief technology officer and chairman of the technical committee of New H3C Group, emphasized the opening strategy many times in an interview with 21st Century Business Herald reporters, "We are open, we can not only use the private domain large model to ensure the security of private domain data, but also support customers to choose the large model combined with New H3C's ICT products, such as New H3C has in-depth cooperation with Baidu." ”

In addition, in terms of the company's overall revenue target, Yu Yingtao, chairman of Unigroup and president and CEO of New H3C Group, said at the Pilot Summit: "The goal of achieving 100 billion revenue by 2025 remains unchanged. ”

The field of AI goes hand in hand

At the table of the AI big model, companies are slotted. From Baidu Wenxin Yiyan, Ali Tongyi Qianqian, Tencent Mixed Yuan Big Model, to iFLYTEK Spark Model, Huawei Pangu Big Model, etc., China is showing a trend of rising together.

The direction of New H3C is a private domain model for B-side, similar to the concept of private cloud.

"Private domain means to deploy within the enterprise and emphasize data security, which is the most critical," You Xuejun said, "On the other hand, as an ICT manufacturer, we also provide AIGC with computing power infrastructure, including network, computing, storage, and algorithm optimization." ”

It can be seen that New H3C wants to provide integrated software and hardware solutions that integrate generative AI, and in addition to the private domain large model, product and technological innovation in the field of servers and switches are also the strengths of New H3C, and this time it has also been upgraded.

According to reports, New H3C's latest UniServer G6 series AI server, equipped with Intel's fourth-generation Xeon Scalable processor and NVIDIA H800 GPU, increased general computing power by 53%, AI computing power increased by 3 times.

Although the overall growth rate of servers has slowed down this year, many institutions are bullish on the demand for AI servers. According to the data provided by TrendForce Consulting to the 21st Century Business Herald reporter, the current AI server mainly equipped with NVIDIA A100, H100, AMD MI300, and large CSP companies such as Google and AWS independently developed ASICs has a relatively strong growth demand, and the shipment of AI servers (including GPUs, FPGAs, ASICs, etc.) in 2023 is estimated to be nearly 1.2 million units, with an annual growth rate of nearly 38%.

At the same time, generative AI is also driving the demand for data center switches. "According to IDC data, the proportion of data center switches will exceed 50% in 2024, and 100G and 400G switches are currently the main force of shipments, and with the large increase in generative AI in 2024, the demand for 800G and 400G will gradually increase." ”

In the industry's view, CPO (photoelectric co-packaging) related technology will become the next technical trend in data centers, and its maintainability and deployable capabilities will continue to be optimized and matured.

Computing power and application challenges

At present, in the process of large model development and generative AI evolution, it faces challenges related to computing power and commercial use.

On the one hand, for the computing power industry chain, GPU demand has skyrocketed, and competition for GPU resources has become an AIGC ticket.

TrendForce Consulting estimates that by 2025, if the world estimates that with 5 ultra-large AIGC products equivalent to ChatGPT, 25 medium-sized AIGC products from Midjourney, and 80 small AIGC products, the computing resources required for the above will be at least 145600~233700 NVIDIA A100 GPUs, plus emerging applications such as supercomputers, 8K video streaming, AR/VR, etc. The load on cloud computing systems will also be increased, and the demand for high-speed computing is increasing.

You Xuejun told reporters: "With the advent of the AIGC era, from the perspective of ICT manufacturers, what we feel the most is that the demand for GPUs is explosive growth. The core of AIGC computing power is still the GPU, and how to optimize the computing power performance on the available GPU and use the least investment to meet the customer's requirements for computing power is a new problem. ”

On the other hand, a number of AI enterprise practitioners told the 21st Century Business Herald reporter that if deployed locally, for enterprise customers, how to reduce the cost of computing power for training and protect data security is also a challenge.

Liu Xinmin, vice president of New H3C Group and president of the technology strategy department, told the 21st Century Business Herald reporter: "The internal data of the enterprise is not so large, and it does not need as much computing power as training a general large model. Opening the internal data of the enterprise to the big model does not need to worry about data security, which is the unique advantage of the private domain big model. But this does not mean that the knowledge of the private domain grand model is barren. ”

In addition, generative AI still faces problems with business models. In You Xuejun's view, whether AIGC can succeed lies in how it combines landing and application, and it is easier to land on the To B side.

You Xuejun said that New H3C's annual R&D expenditure accounts for 10%-15% of sales, and continues to increase investment in cutting-edge innovation projects, which does not talk about short-term returns.

According to the announcement of Unigroup, the HPE entity will sell its 49% stake in New H3C to Unigroup. After the completion of the transaction, Unigroup International will hold 100% of the equity of New H3C. Under the new ownership structure and the new layout of AI, it remains to be seen how New H3C will write a new story.

For more information, please download 21 Finance APP