NVIDIA's H20 chips have caused an uproar in the Chinese market, and despite their low performance and high price, domestic AI companies are rushing to buy them. The story behind this can be described as a "dragon and tiger battle" in the technology world, which not only affects the market, but also reveals the cruel reality of the Sino-US technology game.
Since 2021, the United States has frequently restricted NVIDIA's sales of high-performance AI chips to China, which is simply a draw.
In order to comply with United States export regulations, NVIDIA had to introduce A800 and H800 chips with reduced performance for the Chinese market. In the second half of 2023, the United States will further tighten its policy and simply not even allow the A800 and H800 to be sold.
As a result, NVIDIA rushed to the point of incompetence and launched the H20 chip, which has further shrunk performance, and its performance is only about 15% of that of high-end chips, but its bandwidth is powerful, and powerful performance can be achieved by integrating multiple H20 chips.
Speaking of which, domestic AI chips seem to have ushered in spring in this context. The performance of a domestic AI chip is equivalent to twice that of H20, and the price is similar, which makes domestic AI companies excited.
Since last year, domestic AI companies have tested the waters, replacing NVIDIA's chips with this domestic AI chip, and there is even news that this chip has won 1 million orders.
The ideal is very plump, and the reality is very skinny. Domestic AI companies quickly discovered the problem of domestic AI chips. First of all, the chip area is large and the power consumption is also high, which is simply a "big stomach".
Moreover, when implementing multi-chip integration, the problem of heating is a headache. The bigger trouble is the CUDA ecological problem. The world's AI software is based on NVIDIA's CUDA platform, and the domestic AI chip has been replaced, and the software has to be redesigned, which is simply a redesign.
Despite this, H20 sales in the Chinese market are still booming, selling more than a million copies and bringing at least $12 billion in revenue to NVIDIA.
The proportion of revenue contributed by the Chinese market to NVIDIA has dropped from a quarter to about 9%, which also reflects the changes and challenges in the market environment.
AI companies have encountered more thorny problems in the process of building large models and trying to replace them with domestic AI chips. First of all, it is difficult for domestic AI chips to achieve the high efficiency of NVIDIA chips in actual use, which is a fatal problem for AI companies.
Especially in the case of China's large AI model accelerating to catch up with United States, the time cost is a pressure they can hardly bear.
Tesla's example is a vivid lesson. Tesla has tried to change the direction of its technology and retrain its self-driving technology with large AI models, and found that end-to-end autonomous driving is more efficient.
But they soon realized that the performance of domestic AI chips was not enough to support such efficient operation, and eventually switched back to buying NVIDIA's H20 chips.
This is not only Tesla, but also several domestic Internet companies and autonomous driving companies have followed suit, and since the second quarter of this year, they have significantly increased the purchase of H20 chips.
Foreign media predict that NVIDIA's sales in China will increase significantly this year. Although the performance of the H20 chip has shrunk, its bandwidth advantage and the potential of multi-chip integration still give domestic enterprises hope.
For a certain domestic AI chip, although its nominal performance is very strong, the actual performance is disappointing. After the initial attempt, the final purchase volume may only be 500,000 sheets, which is only about half of what was originally expected.
It is undeniable that although the performance of domestic AI chips is not satisfactory, it is still a huge progress. The order of 500,000 AI chips still brought it more than $6 billion in revenue.
Compared with the previous monopoly of ninety percent of the domestic AI chip market by NVIDIA, this is a huge breakthrough. More importantly, through these attempts, domestic AI companies have discovered problems and promoted the technological upgrading of domestic AI chips.
I believe that the prospect of domestic AI chips replacing NVIDIA chips is still bright. Although there are many challenges at present, the continuous progress of technology and the demand of the market will push domestic chips to go further and further. It's like climbing a mountain, even though the trail is rough, if you persevere, you can see the scenery at the top.
(Disclaimer) The process and pictures described in the article are from the Internet, and this article aims to advocate positive social energy and no vulgar and other bad guidance. If it involves copyright or character infringement issues, please contact us in time, and we will delete the content as soon as possible! If there is any doubt about the incident, it will be deleted or changed immediately after contact.