Original Source: Dongbu Yaqiao Research Institute
Image Source: Generated by Wujie AI
NVIDIA has launched a cut-down version of GPU chips, H20, L20, and L2, specifically for the Chinese market, based on NVIDIA's Hopper and Ada architectures. It is said that mass production will take place around December this year to January next year.
Similar to A800 and H800, NVIDIA once again targets the red line of the US 1017 export control, achieving compliance with the new US chip export control by restricting certain parameters.
The updated export control rules by the US Department of Commerce use "total processing performance" and "performance density" as the criteria for controlled chips, replacing the previous standards based on peak computing power and bandwidth, and no longer restricting bandwidth. This change brought NVIDIA's A800 and H800, which were specifically for China, into the controlled range after the 1007 rules last year, and it took effect immediately. These chips that NVIDIA has not yet released will immediately be unavailable to China.
L20, L2, and advanced AI large model training are not closely related, mainly playing a role in inference and edge computing. H20 is directly related to large model training. The technical details are not elaborated here, as SemiAnalysis has already provided specific analysis. In simple terms, H20 has high cache, high bandwidth, and low computing power. NVIDIA ensures that H20 does not exceed the "total processing performance" and "performance density" in the 1017 rules by limiting the peak computing power of the H20 GPU chip, thereby making it not a controlled chip. At the same time, NVIDIA has increased the bandwidth of H20 (because the 1017 new rules no longer restrict bandwidth) and storage capacity (1017 also does not have requirements for cache).
The speed and efficiency of large model training depend on many factors, such as GPU computing power, bandwidth, memory (VRAM) capacity, storage I/O performance, data loading and preprocessing, model architecture, software architecture, algorithms, network latency, and heat dissipation. Computing power and bandwidth are particularly important. With high GPU computing power, parallel computing operations can be significantly accelerated, processing larger datasets in a reasonable time, accelerating model iteration and optimization, having more computing resources to handle more parameters and layers, and implementing more complex model training. Sufficient bandwidth ensures that data is quickly transferred to the GPU for parallel processing, ensuring that data transfer speed can keep up with GPU processing speed, fully utilizing the GPU's processing power, shortening model training time, and enabling faster model updates.
The bandwidth of A100 is 600G/s, while A800 and H800 are both 400G/s. H20 uses NVIDIA's unique NVLink interconnect technology to increase the bandwidth to 900G/s (the same as H100). Because NVIDIA has retained the complete NVLink and PCIe bandwidth functions, it means that there is no problem with connecting tens of thousands of cards for parallel computing. Although the total processing performance of a single H20 card is not as good as H100, due to sufficient bandwidth, multiple cards can be stacked, and multiple AI servers can be clustered together to compensate for the insufficient computing power of a single card. At the same time, H20's storage capacity has increased by 16G compared to H100. Sufficient storage capacity is also very important for storing the parameters and intermediate data of large models. Insufficient memory may require processing data or models in batches, thereby reducing the efficiency of large model training.
The timing of NVIDIA's launch of H20 is also intriguing. Huawei's Ascend 910B's performance is very close to the overall performance of A100 single card after the 1017 new rules. Therefore, everyone thought that Huawei might become the only GPU provider in the domestic market. It has also been reported these days that Baidu and 360 have made large purchases of Ascend 910B. However, compared to NVIDIA GPU chips, Ascend 910B has the disadvantage of lower interconnect bandwidth and less mature software ecosystem. The launch of H20 by NVIDIA is precisely based on its own CUDA software ecosystem, significantly increasing the interconnect bandwidth, which is quite obvious in comparison to Huawei. As a result, domestic large model manufacturers may not only consider Huawei's Ascend series, and many companies may choose to buy H20 after weighing the options, which is clearly not favorable for Huawei. However, for NVIDIA, it has once again bypassed US export controls and can sell more chips (many Chinese AI companies may buy more H20 to achieve the performance of A100 through clustered clusters).
How do the US and Chinese governments view this? They may both be a bit conflicted. The US Department of Commerce should be very unhappy with NVIDIA over the past two years because it has obviously repeatedly violated the spirit of US chip export controls, helping Chinese companies evade US export controls. From the fact that the Department of Commerce "specially treated" NVIDIA after the 1017 new rules were announced and immediately notified it of the immediate effect of the new rules, it seems that they are also aware of this. Raymond has said on several occasions that although chip control will cause some losses to US companies, it is worth it for national security. US companies cannot just focus on making money without taking responsibility for national security. These words seem to be specifically for Jenson Huang.
However, the US Department of Commerce may be more helpless about NVIDIA's actions. Currently, the Department of Commerce can only use export controls as a tool, but regulation and law are always behind technology, business, and technological innovation. NVIDIA's quick introduction of a new cut-down version of the chip clearly illustrates this issue. In this situation, what will the Department of Commerce do next? Will it wait until next year to modify the export control rules and reintroduce bandwidth? Will NVIDIA find another way to circumvent it? This cat-and-mouse game should make Raymond very frustrated and disturbing, but for the time being, there seems to be no solution, after all, NVIDIA is acting within the framework of US law. The weakness of export controls in chip export controls is fully exposed. Therefore, Raymond said during a congressional hearing on chip law enforcement that the Department of Commerce "needs new tools." Whether it is to deal with Huawei or NVIDIA, she really needs more effective tools than export controls.
However, for the United States, which advocates economic and trade liberalization and a free market economy, direct government intervention in US companies and foreign business transactions is not in line with its own values. If export controls cannot prevent NVIDIA from selling high-end GPU chips to China, and the US government hopes to completely block NVIDIA's evasion, it may have to give up starting from "chips" and instead find ways to block transactions between NVIDIA and Chinese companies. With my limited knowledge of US sanctions and national security laws, I can only think of the International Emergency Economic Powers Act (IEEPA). If friends who understand US law can think of other options, please feel free to share.
The IEEPA authorizes the US President to take measures to control international business and economic transactions when declaring a national emergency, including freezing assets, restricting trade, and prohibiting financial transactions, but this requires the cooperation of US companies to be effectively implemented. A possible scenario is that the United States believes that allowing Chinese AI companies to continue to obtain advanced GPU chips for training large models poses a serious threat to US national security. In theory, the President can declare a national emergency based on the IEEPA and then issue an executive order requiring US GPU chip companies not to engage in any transactions with Chinese companies.
Ironically, the IEEPA was originally intended to limit the President's power to restrict the economy under the Trading with the Enemy Act (TWEA) during World War I, but it later became a common legal tool for the United States to impose sanctions and prohibit US companies from trading with sanctioned countries. However, the IEEPA was previously used in small countries such as North Korea, Venezuela, Iran, and Central Africa, where US companies had relatively weak commercial ties with these countries, so they cooperated well with the US government. However, if it is applied to China, a country with a large number of commercial ties with the United States, the situation will naturally be very different, and the backlash from US companies will be stronger.
On the flip side, after the 1017 new rules, it was quickly reported that Huawei was able to develop AI chips that are comparable to NVIDIA's GPU chips. Some US think tanks also questioned whether the 1017 new rules could restrict Chinese companies, and instead helped Huawei a lot. Although NVIDIA's H20 evades US export controls, it invisibly diverts many Chinese customers from Huawei's Ascend 910B, preventing Huawei from dominating the Chinese market, which seems to be what the US Department of Commerce would like to see.
It is also worth observing China's reaction. The strategic direction of the US chip export control is computing power and AI, focusing on the strategic competition between China and the United States in the emerging technology and key industry of AI. The Chinese government and policy circles should have a clear view of this, so the determination for domestic substitution is very firm. Many opinions may even believe that the current constraints are due to the lack of firm and early development of domestic products. China does not hold much hope for the relaxation of chip control by the United States. The current strategy seems to be to use export controls on key minerals to a certain extent as a countermeasure, while also showing signs of gradually cutting off bilateral issues with the United States, that is, technological competition is only about competition, but does not affect political agendas involving the overall Sino-US relationship.
Huawei's deep involvement in the chip field has injected the efficiency of private enterprises and higher execution and integration capabilities into China's semiconductor industry from design to manufacturing. This includes a series of visible achievements, including the 7nm mobile chip, which has greatly increased the confidence of the Chinese government and industry in pursuing domestic substitution. Innovations in RISC-V, 3D stacked chip technology, nano-imprinting, and other technologies may also enable China to take a different technological path and achieve success.
In this situation, the most ideal scenario is definitely the formation of a more cohesive chip industry ecosystem centered around Huawei domestically, and it is not desirable for NVIDIA to disrupt this. Huawei's rotating chairman Xu Zhijun recently called on domestic companies to give up their illusions and use and support domestic chips on a large scale, even if they are difficult to use, to help the Chinese chip industry quickly break free from US control and achieve independence and supply chain security through continuous adjustment. However, Huawei's GPU chip production capacity is still limited, and it is difficult to meet the needs of domestic large model training in the short term. If NVIDIA can provide alternatives such as H20, it will definitely be a positive development for the domestic AI industry and will help China alleviate the progress lag caused by chip shortages in this round of AI competition.
In conclusion, the game around this small GPU chip is becoming increasingly lively.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
