In recent years, the United States has continuously escalated the export controls on AI chips to China. Even NVIDIA, the leading AI chip manufacturer, has faced repeated restrictions on its "Special Edition" products tailored for the Chinese market.
As early as October 2022, the United States officially introduced policies restricting the export of AI chips to China. Under ECCN 3A090 and 4A090 categories, high-performance computing devices, chips, and components meeting the following two conditions are considered controlled high-performance computing chips:
This move directly led to NVIDIA's high-performance AI chips like A100 and H100 being unable to be exported to China. Subsequently, in an effort to salvage the Chinese market, NVIDIA introduced AI chips tailored for the Chinese market that complied with U.S. restriction policies: A100 and A800, mainly by reducing the interconnect speed.
However, on October 17, 2023, the U.S. government introduced new restriction rules, further tightening the scope of restrictions:
Any U.S. chips falling within the performance standards mentioned above will be restricted. This has also directly led to restrictions on the export of NVIDIA's A100 and A800 chips tailored for the Chinese market. NVIDIA has incurred significant losses in its investment in these two chips. Additionally, NVIDIA's L40S and high-end graphics card RTX 4090 are also subject to restrictions.
However, NVIDIA is still unwilling to give up on the vast AI chip market in China. Therefore, in the last quarter of last year, NVIDIA developed four AI chips tailored for the Chinese market that comply with the latest U.S. restriction policies, including HGX H20, L20 PCle, and L2 PCle. They also developed the high-end graphics card RTX 4090 D, which complies with the new regulations.
Based on the specs that have been released, the top-of-the-line H20 model's main computational power parameters such as FP16 and INT8 are less than half of those of the A100, and only about one-seventh of the H100's. The main computational power parameters of the L20 have also decreased by about one-third compared to the previous L40 and two-thirds compared to the L40S. In an interview, an insider from NVIDIA also admitted that the H20's single-card computing power is only 20% of the H100's, and compared to a certain domestic chip, its performance is just over 60%.
Moreover, the core count of the RTX 4090 D has been reduced by about 11% compared to the RTX 4090, potentially decreasing the overall performance by 10%!
Clearly, NVIDIA's latest products customized for the Chinese market have been significantly downgraded, leading to a generally pessimistic or skeptical attitude towards their performance and cost-effectiveness in the market.
Despite this, it seems the United States still has concerns.
Recently, the U.S. announced new restriction rules set to take effect on April 4th of this year. Under the category 4A003, "digital computers," "electronic assemblies" and related equipment and "components" destined for all countries (except those listed in Supplement No. 1 to Part 740 of the EAR, countries in groups E:1 or E:2) that have an "Adjusted Peak Performance" (APP) exceeding 70 Weighted TeraFLOPS (WT, trillions of floating-point operations per second, primarily for GPU performance, with CPUs and NPUs mainly providing fixed-point/integer computing power) will require a license (NLR).
It's worth mentioning that in AI training, the floating-point formats FP16 and FP32 are commonly used due to their high precision. However, for AI inference, integer data formats like INT8 and INT4 are typically employed. This means that for U.S. manufacturers' computing products or components with an aggregate floating-point computational capability exceeding 70 TFLOPS, an export license is required. This seems to target Nvidia's newly introduced RTX 4090-D and H20 for the Chinese market.
Data indicates that the RTX 4090-D has an FP16/FP32 performance of 74TFLOPS, and the H20 has a TF32 performance of 74 TFLOPS (with the L20's FP32/16 performance being just under 60TFLOPS). This implies that Nvidia's RTX 4090-D and H20 may require a license for export to China, and whether approval will be granted might need to be determined on a "case-by-case review".
Although some domestic media reports suggest that this policy could restrict high-performance AI PC products, Chip Intelligence believes that the floating-point computational power of processors aimed at AI PCs from manufacturers like AMD, Intel, Qualcomm, and Apple, which integrate AI cores, is still relatively limited. They are more focused on enhancing integer computational capabilities and are unlikely to be affected by the new regulations. For example, Intel's Meteor Lake chip (Core Ultra), launched at the end of last year, based on its AI engine, NPU, CPU, and GPU, can deliver 34TOPS of performance (note that's TOPS, not FLOPS).
However, this is just the author's interpretation of the rules. So far, Nvidia has not released any official information on this matter. Chip Intelligence has reached out to Nvidia's representatives in China, but they have not responded.
It's noteworthy that U.S. Commerce Secretary Gina Raimondo, during the Reagan National Defense Forum held on December 2nd, stated that despite the difficulties it might pose for businesses, the U.S. government will continue to constantly change the export controls on advanced chips because "technology changes, adversaries change, and we have to keep up". Raimondo pointed out that if a company redesigns a chip to circumvent current regulations, but it's used by a country of concern to develop AI, "I’m going to control it the very next day".
Responding to Raimondo's comments, Nvidia CEO Jensen Huang said during a press conference in Singapore on December 6th that Nvidia has always been actively cooperating with the U.S. government to create products that comply with regulations. "We plan to continue cooperating with the U.S. government and develop a new series of products that comply with the new regulations". Huang had previously stated that U.S. restrictions on semiconductor exports to China have further motivated China to develop its semiconductor industry. Currently, dozens of companies in mainland China are developing technologies to compete with Nvidia's products, which is not favorable for Nvidia's development in the Chinese market. Over the past years, the Chinese market has accounted for about 20% of Nvidia's revenue.
