Nvidia H20 is expected to arrive in March, Chinese companies wait and see and do not "pay"
The performance of the "China Special Edition" has shrunk severely
Author/ IT Times reporter Jia Tianrong
Editor/ Hao Junhui Sun Yan
Recently, according to Reuters, Nvidia plans to mass-produce the H20 AI chip designed for China in the second quarter to comply with U.S. export regulations.
The report said that H20 was supposed to be launched in November 2023, but server manufacturers have encountered difficulties in integrating the chip, and H20 will prioritize meeting the needs of large Chinese customers.
The "IT Times" reporter learned that some domestic manufacturers will get H20 in the first quarter of this year, but due to the significant shrinkage of the performance of these "special version" AI chips, most Chinese companies are still in a wait-and-see state.
It will arrive in the first quarter
"China Special Edition" may be cold
According to industry industry sources, Nvidia is developing the "latest improved" AI chip for China, and the three chips are all based on Nvidia H100 to comply with the latest technology export control policies of the United States, including HGX H20, L20 PCle and L2 PCle.
In view of the prospect of special chips in the Chinese market, industry insiders revealed to the "IT Times" reporter that domestic partners have purchased Nvidia H20 chips, which can be the first batch to arrive in March, but whether to purchase in large quantities, "then it depends on the effect of the test."
Due to the serious shrinkage of performance, most manufacturers are not optimistic about the special version of the H20.
According to the previously exposed parameter information, Nvidia H20 is in the same series as H100 and H200, all of which use Nvidia Hopper architecture, but the memory capacity has been increased to 96GB HBM3, and the GPU memory bandwidth is 4.0TB/s. In terms of computing power, the FP8 computing power of this product is 296 TFLOPS, and the FP16 computing power is 148 TFLOPS, which is 80% smaller than that of H100 and is only 1/13 of the "strongest" AI chip H200.
"As far as large model training is concerned, H20 is basically an unusable machine. Peng Lu, COO of Shanhai Engine, told the IT Times that the configuration of H20 is more suitable for inference than for model training. Peng Lu believes that from the perspective of market application, except for some large manufacturers that may invest in inference, large model startups rarely buy inference computing power in large quantities, because they pay more attention to the competitiveness of model training.
In fact, the big manufacturers are not satisfied with the results of the H20 sample test. According to people familiar with the matter, Alibaba Group, Tencent and other companies have been testing Nvidia's special chip samples since November 2023, and have indicated that the number of chips ordered from Nvidia this year will be far less than the banned Nvidia high-performance chips that were originally planned to be purchased. It was previously reported that in 2023, Chinese internet companies placed $5 billion in orders for AI chips from Nvidia.
Borrow car chips
Push up the moat
China is one of the most important markets for Nvidia, and Nvidia pointed out in its earnings report that China and some other regions affected by restrictions bring about a quarter of its data center revenue, of which the Chinese market accounts for the vast majority.
For Nvidia, it is challenging to meet the requirements of U.S. regulators and the Chinese market, and it is focused on expanding the advantages of consumer GPUs such as gaming and autonomous driving.
At the end of 2023, NVIDIA officially launched the GeForce RTX 4090D, which is the first chip focused on the Chinese market officially launched by NVIDIA since the Biden administration announced the new chip export rules in October 2023. In the early morning of January 9, 2024, Nvidia released three RTX 40 SUPER series graphics card products for consumers at its pre-CES event, namely RTX 4080 SUPER, RTX 4070 Ti SUPER and RTX 4070 SUPER, all three products are based on TSMC's 4nm process technology and use the new Ada Lovelace architecture. Nvidia said that the three new graphics cards will comply with export controls and can be sold in China.
GeForce RTX 4090D
At CES 2024, NVIDIA announced cooperation with a number of Chinese companies in the field of games and autonomous driving, including miHoYo, NetEase Games, PalmFun Technology, Tencent Games, etc., to cooperate with them in the field of AI technology application and apply it to game development.
In addition, Li Auto will use the Thor automotive chip platform for next-generation models. At present, Great Wall Motor, ZEEKR and Xiaomi Auto have adopted Orin chips to create a new generation of intelligent driving systems. According to the data of the Gaogong Intelligent Vehicle Research Institute, as of the first half of 2023, NVIDIA's market share in models with high-level assisted driving functions (NOA) has reached 52.57%.
These moves have also brought confidence to investors. On January 11, Nvidia hit record intraday and closing all-time highs for the third consecutive trading day. Brokerage Truist raised its price target on Nvidia to $691 from $674.
Colette Kress, Nvidia's chief financial officer, said that under U.S. regulations, certain export products are licensed, and the company is working with customers in those regions to try to provide a "solution" that is licensed to ship products and does not trigger U.S. government restrictions. She even said that without the new rules for the shipment of AI chips in China, Nvidia's performance outlook for the fourth fiscal quarter (ending January this year) will be higher.
Is there an opportunity for domestic computing power?
Analysts at TrendForce, a technology research company, pointed out that about 80% of the high-end AI chips of Chinese cloud computing companies come from Nvidia, and this proportion may drop to 50%~60% in the next five years, which may provide opportunities for domestic chips to overtake in corners.
However, it is undeniable that there is still a gap between domestic AI chips and NVIDIA's top products. A number of industry insiders told the "IT Times" reporter that Nvidia is still the largest chip provider in the domestic market, and the servers equipped with Nvidia H800 or A100 chips that appear sporadically in the market are still in short supply.
NVIDIA's ecosystem is a huge advantage, and if it switches to domestic chip servers, enterprises need to redevelop and adapt the original training data and solve a series of technical challenges.
A telecom operator told the "IT Times" reporter that when choosing a chip, the factors considered include price, computing performance, purchase cost, sales price and maintenance cost. At present, Huawei's Ascend 910B is the closest product to NVIDIA's top chip in China, but the industry consensus is that the production capacity of 910B is still not ideal.
It is reported that the shipment of Ascend 910B this year is more than 400,000 units, and the price is still rising. Therefore, some powerful companies choose to "bet on both ends".
In terms of opportunities in China's chip field, Peng Lu believes that there are two main factors. The first is ecosystem support, which enables general-purpose GPUs to reduce the adjustment cost of adaptation for developers, and the second is privatized deployment, which is oriented to the development of industry solutions and industry applications.
Peng Lu predicts that this field may usher in a big explosion in the next two years, "Many enterprises, especially state-owned enterprises, are more inclined to choose domestic chips to promote the development of the entire domestic chip industry in the market." ”
Typesetting / Ji Jiaying
Image / Nvidia Touhou IC
来源／《IT时报》公众号vittimes
E N D