Against the backdrop of global computing resource scarcity, NVIDIA, the "military leader" of AI chips, has accelerated its product iteration speed and recently launched the high-end artificial intelligence GPU chip H200. It is expected to start supply in the second quarter of 2024 and drive the high-end memory market next year.
From the perspective of market performance, on November 14th, the A-share artificial intelligence sector significantly boosted. After Huina Technology announced the price increase of computing services, the stock price quickly rose, driving the computing leasing sector higher. A listed company with computing power service provider confirmed to a reporter from Securities Times · E Company that the company has discussed price increases with downstream customers.
Accelerate upgrades and increase memory
Whether it's generative AI or high-performance computing, the market's demand for high-capacity and ultra high-speed GPUs is increasing rapidly. Nvidia AI chips have become a "hard demand", shifting the pace of architecture releases over the past two years to annual updates and accelerating upgrades to respond to market demand.
On November 13th local time, Nvidia announced the release of a new generation H200 chip based on the AI chip H100. The testing of the Llama 2 model based on Meta shows that the output speed of H200 is approximately twice that of H100. In addition, the inference speed of H200 is almost twice that of H100, and the bandwidth has increased by 2.4 times.
Compared to the previous generation A100 and H100, the main change of H200 lies in its memory, becoming the first GPU to use HBM3e (High Bandwidth Memory), increasing the bandwidth from 3.35TB per second in H100 to 4.8TB, an increase of 1.4 times. The total memory capacity has increased from 80GB in H100 to 141GB, an increase of 1.8 times, and the inference energy consumption has also been significantly reduced.
According to Trend Force's research, the current mainstream in the HBM market is HBM2e, including NVIDIA A100/A800, AMD MI200, and most cloud service providers' self-developed acceleration chips designed to this specification. At the same time, in order to comply with the evolving demand for AI accelerator chips, each original factory plans to launch a new product HBM3e in 2024, and it is expected that HBM3 and HBM3e will become the mainstream in the market next year. Recently, major storage companies such as Samsung and Micron are continuously accelerating their expansion of production.
According to the latest prediction by Wu Yating, an analyst at Jibang Consulting, the global total production of HBM is expected to double and become a major contributor to the revenue of DRAM major factories in 2024.
Domestic storage manufacturers mostly focus on enterprise level storage products and layout the computing server market. Baiwei Storage recently stated on the interactive platform that the company's enterprise level SSDs and server memory modules can be applied to computing servers. Currently, the shipment of these products accounts for a small proportion of the company's total revenue. The company's high-end storage chips include ePOP storage chips for flagship wearable products such as smartwatches and AR/VR, as well as UFS3.1, LPDDR5, uMCP storage chips for mid to high-end mobile phones.
During a recent survey, executives from Jiangbolong stated that in the company's enterprise level storage product portfolio, eSSD products and RDIMM products are mainly used in high-end enterprise level application scenarios mainly focused on servers. At present, the company's eSSD and RDIMM products have been certified by important customers such as Lenovo, JD Cloud, and BiliBili, and have obtained official orders from some customers to achieve mass production and shipment, which will bring new growth to the company's performance.
Giants compete in the Chinese AI chip market
The H200 launched this time will be launched in the second quarter of 2024, and Nvidia's CoreWeave, Lambda, and Vultr investments, as well as Amazon Cloud Technology, Google Cloud, Microsoft Azure, and Oracle Cloud infrastructure, will all become H200 "new" users.
Around the Chinese market, computing power suppliers are facing increasingly strict controls. On October 17th of this year, the US Department of Commerce's Bureau of Industry and Security (BIS) released new regulations on chip export controls, setting "performance density thresholds" and other indicators, and upgrading the control of AI chips with high computing power; And it quickly tightened the 30 day policy grace period. On October 24th, Nvidia announced that its data center or sales of products with a total computing power of 4800 or higher were required to immediately implement regulatory measures.
The giants are also adjusting their product strategies around US government regulation. Nvidia has previously provided low-end versions of the 800 series AI chips to the Chinese market. After regulatory upgrades, media reports have recently reported that Nvidia has developed the latest improved series of chips HGX H20, L20 PCle, and L2 PCle for the Chinese market. However, Nvidia has not yet made any official comments on this. In September of this year, Intel announced that it would launch the high-end AI processor Gaudi 3 using a 5nm process next year, and previously launched the high-end AI processor Gaudi 2 for the Chinese market.
A listed computing power operator told a reporter from E Company that it has noticed that Nvidia has provided a special supply of computing power chips to the Chinese market, and whether it will purchase them is subject to official information.
In addition, the staff of the First Financial Reporting Inspur Information stated that they have not received any relevant notification yet, and the company will continue to communicate with the other party; If the message is true and the customer's acceptance is good, there may be some order delivery, but the specific situation still needs to wait for the authenticity of the message. The Secretary General's Office of Ukede responded to the media that it is considering procurement and will consider the adaptation effect after the manufacturer provides the chip.
Against the backdrop of insufficient supply and demand for NVIDIA AI chips, the demand for AI computing power leasing, construction, and other services has exploded. According to incomplete statistics from E Company, approximately 10 listed companies have recently issued and signed agreements related to computing power services.
Computing Power Market Initiates Price Increase
At the time of NVIDIA's new product launch, a strong "price increase" model was launched in the computing power market due to insufficient supply. Huina Technology announced at the close of the morning on November 14th that it plans to increase the pricing of some computing services.
According to the notice from the partner Sichuan Biji Technology Co., Ltd., due to the significant increase in computing power demand for high-performance computing servers embedded with NVIDIA A100 chips, the prices of related high-performance computing equipment continue to rise, and computing resources continue to be scarce. Sichuan Biji Technology Co., Ltd. has decided to raise its A100 computing power service fees by 100%. Starting from today, Huina Technology plans to simultaneously increase the computing power service fees of the high-performance computing power server embedded with NVIDIA A100 chips that it is entrusted to operate by 100%.
Although Huina Technology stated that the price adjustment will not have a significant impact on business performance for the time being, the company's stock price began to rise before the announcement was released, rising nearly 20% from 10:53, and the stock price was closed to the limit up board.
According to a reporter from E Company, the supply of computing power in the market has become increasingly tight recently, and even non high-performance computing power has encountered some difficulty in purchasing. Recently, several companies have reported that computing power costs have increased by about 30% -50%, and the price increase of Huina Technology has once again raised the industry's water level. However, Huina Technology is targeting the A100 series, which currently has the strongest computing power.
The staff of Zhongbei Communication's Secretary Office told the reporter of E Company that the company is communicating with customers about price increases. In addition, the company's ongoing computing power center is waiting for equipment delivery, and the AI chips purchased include the Nvidia A/H 800 series. It is currently uncertain whether there will be a delay, and the final delivery time will prevail.
In November, during an institutional survey, a senior executive at Zhongbei Communication revealed that the supply of computing power has been relatively tight recently, and many channels are currently unable to obtain goods. Therefore, the corresponding prices will gradually increase, and negotiations are underway with customers to discuss price increases. It is reported that the company plans to establish four intelligent computing centers and expand the overseas computing service market. In recent years, the signed amount of orders has reached 865 million yuan, and it is expected that the total signed amount for the whole year will exceed last year.
In addition, according to media reports, staff members from the Investor Relations Department of Hengrun Corporation stated that the procurement cost has increased by 50%, and the rental price of computing power is also increasing; Qingyun Technology stated on the interactive platform that the company will increase prices in response to recent changes in market supply.
Faced with the current tight supply of computing power chips, domestic computing power operators have teamed up with domestic GPU manufacturers to launch a "combination punch" in computing power supply.
Zhongbei Communication executives stated that while further expanding the procurement channels and efforts of internationally leading performance servers, the company has strengthened communication and cooperation with domestic GPU manufacturers. Recently, we have had exchanges with Huawei regarding cooperation in the Huawei Root Technology Ecological Alliance, technology and business cooperation, and have reached a computing power cooperation intention with Hyperfusion. The deployment plan for domestic computing power is being refined and implemented.
Hongxin Electronics stated on the interactive platform that the company has maintained close contact with NVIDIA and related channel providers, has a good cooperative relationship, and currently has sufficient procurement channels for NVIDIA's multiple model chips. In the delivery of the company's upcoming 1 billion server order, in response to the market demand of the end customer, the company will provide a combination of hybrid computing power, flexible configuration of NVIDIA computing power and domestic Suyuan computing power, among which NVIDIA computing power is the main one to meet customer needs to the maximum extent.
On the evening of October 24th, Hongxin Electronics announced that the company has signed a one-year "Purchase Agreement" with Shanghai Suiyuan Technology Co., Ltd. The company will flexibly configure the hybrid computing power of the Suihong brand according to the actual needs of end customers. This year, the servers to be delivered will be partially equipped with domestic computing power such as Suiyuan chips, and partially equipped with imported computing power such as Nvidia chips, ultimately forming a diversified heterogeneous hybrid computing power; At present, the company has officially issued a purchase order for 9152 chips to Suiyuan Technology, and plans to complete the delivery of all chips by the end of the year, quickly forming a domestic computing power landing.