The China-specific version of the Gaudi 3 needs to significantly reduce AI performance in order to be exported. Text: Technode
Author|Huang Chen typesetting|Li Tingting This article is expected to read for 8 minutes Recently, some foreign media said that Intel is ready to launch a "special version" Gaudi 3 for the Chinese market, including two hardware forms: OAM compatible mezzanine card named HL-328 and PCle accelerator card named HL-388. According to the report, Intel disclosed the above information in its Gaudi 3 white paper, of which HL-328 will be launched on June 24 and HL-388 will be launched on September 24. Shockingly, based on the number of cores, operating frequency, TDP and other parameters, compared with the Gaudi 3 international version, the performance of the Chinese "special version" HL-328 chip may be reduced by about 92%. In terms of specific hardware specifications, the Gaudi 3 of the China Special Edition has the same 96MB of SRAM on-chip storage, 128GB of HBM2e high-bandwidth storage with a bandwidth of 3.7TB/s, and has a PCIe 5.0 x16 interface and decoding standard compared to the original. However, due to the restrictions on the export control rules of AI chips in the United States, the overall computing performance (TPP) of this high-performance AI needs to be less than 4800 before it can be exported to China, which means that the 16-bit performance of the China-specific version of the Gaudi 3 cannot exceed 150 TFLOPS. According to the information released by Intel, Gaudi 3 can reach 1835 TFLOPS on FP16/BF16, which is 40% faster than NVIDIA H100 in large model training and 50% more energy efficient in inference. Obviously, the China-specific version of Gaudi 3 needs to significantly reduce AI performance in order to be exported. As a result, the Gaudi 3 for the China Edition needed to drastically reduce the number of cores (the original version had 8 matrix math engines and 64 tensor cores) and operating frequency.
In July last year, Intel released the Gaudi 2 for the Chinese market. Compared to the international version of Gaudi 2, the accelerator card launched for the Chinese market is not much different in performance, and the number of integrated Ethernet RDMA ports has been reduced from 24 to 21 to comply with US chip export control regulations. How does the United States hijack computing power In the 90s of the 20th century, the United States accounted for more than one-third of global chip production, and this share has dropped to about 12% by 2020. In order to maintain its leading position in the semiconductor field, since the U.S. issued the CHIPS and Science Act (hereinafter referred to as the "CHIPS Act") in August 2022, the U.S. has implemented comprehensive semiconductor export controls to China, and the restrictions have been escalating from the chips themselves to the chip manufacturing equipment. The CHIPS Act is at the heart of the Biden administration's revival of industrial policy, which uses U.S. government funds to restore domestic production of technology components critical to national security and economic growth. The bill prohibits subsidized U.S. and its allies from building or expanding advanced process chip fabs in China and other countries of concern for a decade. In October 2022 and October 2023, the Bureau of Industry and Security (BIS) of the U.S. Department of Commerce issued export controls on China's advanced semiconductors and computing equipment twice in a row, in an attempt to affect China's advanced manufacturing, and many GPU and AI chip products from NVIDIA, AMD, AND Intel can no longer be exported to china, and even the high-end gaming graphics card rtx 4090 HAS BEEN RESTRICTED. In December 2023, the U.S. Department of Commerce's BIS announced the launch of an investigation into the semiconductor supply chain of mature process nodes, which is even more clearly aimed at China's chip semiconductor industry. In the early morning of March 30, Beijing time, the Bureau of Industry and Security (BIS) under the U.S. Department of Commerce issued new regulations to "implement additional export controls", revising the two new export restrictions formulated by BIS in October 2022 and October 2023, comprehensively restricting the sales of NVIDIA, AMD and more advanced AI chips and semiconductor equipment to China. In the new regulations, the stick of sanctions is waved again. BIS has removed and revised some restrictions on the sale of semiconductor products to China in the United States, Macau, China, and other places, including the "presumptive refusal policy" for Macau, China and the D:5 country group, and the "case-by-case review" policy rules for AI semiconductor products exported by the United States to China, including comprehensive verification of information such as technical level, customer identity, and compliance program.
Intel's Courage Comes From Although it is not yet available, Intel's special edition of Gaudi 3 is highly likely to pose some potential problems. For example, the reduced performance may affect the user experience and application effectiveness of Chinese enterprises, and at the same time, if the price advantage of the special version of the chip is not competitive, then its market competitiveness may be affected to a certain extent. As a result, Intel needs to make reasonable trade-offs in terms of product design and pricing. Two months ago, Nvidia's "special version" of the AI chip H20 was available for pre-order. The product form includes a computing card and a server equipped with 8 H20 computing cards. In terms of performance, the performance of the Nvidia H20 is about one-sixth of that of the H100, but the price has not been significantly reduced, so the cost performance is not high. At the beginning of this year, according to people familiar with the matter, since November last year, Alibaba, Tencent and other large Chinese companies have been testing Nvidia's special chip samples. They have indicated to Nvidia that the number of chips ordered from Nvidia this year will be far less than the banned Nvidia high-performance chips that they had planned to buy. Even in the face of the risk of declining revenue, Intel is still doing well under the "careful budget". Nearly two years after the introduction of the U.S. government's "CHIPS Act", the veteran chip giant Intel announced in March that it had received up to $8.5 billion in government subsidies and up to $11 billion in special loan support. It is understood that the subsidy support received by Intel comes from the "Chips Act" introduced by the Biden administration in 2022, which strives to help chip companies build more chip factories in the United States and build the United States into a chip manufacturing power. From the perspective of the AI market, Nvidia currently occupies an absolute advantage in the chip market, and it is not easy for Intel to use products to pry away the share. According to Wells Fargo statistics, Nvidia currently has a 98% market share in the data center AI market, while AMD's market share is only 1.2%, and Intel's market share is less than 1%. Therefore, for Intel, it is a wise move to keep up with the US government.
With the rapid development of the digital economy, especially the outbreak of AI, the demand for computing power in the whole society is growing rapidly. According to the "2023-2024 Chinese Intelligent Computing Power Development Assessment Report" jointly launched by IDC and Inspur Information, during the 2022-2027 period, the compound annual growth rate of China's intelligent computing power is expected to reach 33.9%, and the scale of intelligent computing power will reach 1117.4 EFLOPS by 2027. At the same time, the staff of the Southern Branch of the China Academy of Information and Communications Technology said at CITE 2024 that at present, the mainland's intelligent computing power accounts for more than 30% of the world, mainly relying on Nvidia GPU chips in the United States, and the share of domestic independent computing power is only 5%, and the usage rate of American AI frameworks such as Tensorfiow PyTorch and Caffe in China exceeds 90%. From the perspective of application, at present, domestic mainstream chip manufacturers such as Ascend, Cambrian, and Tiantian Zhixin have completed the adaptation of mainstream large models. Industry analysts believe that although there is still a big gap compared with Nvidia and AMD's advanced chips, domestic GPU chips such as the Ascend 910 series can basically support domestic large-scale model applications, and Liu Qingfeng, chairman of iFLYTEK, said at last year's 1024 Developer Festival that Huawei's GPU capabilities have been able to benchmark against Nvidia A100, and based on the Ascend ecosystem, the "Feixing No. 1" large-scale model computing platform has been launched. Previously, the Cambrian Jisyuan (MLU) series of cloud intelligent acceleration cards and the "Zhixiang multi-modal large model" self-developed by Zhixiang Future have also been adapted, and it claims to have reached the level of international mainstream products in terms of product performance and image quality. China's large-scale substitution of imported AI chips is accelerating. For Intel, the key is how to meet the requirements of U.S. policies while taking into account the needs of the Chinese market, maintaining product competitiveness and key customer experience. On the other hand, it also provides valuable opportunities for the development of China's local AI chip manufacturers, who need to pay close attention to market dynamics and technology development trends to cope with potential competitive pressures.
This article is an original article by TechNode reporters, and may not be reprinted without authorization.
--------THAN--------
Interactive topics
What do you think of the Gaudi 3 for the China Special?
Come and leave your opinion in the comment section!
*If you want to get industry information and share your experience with like-minded technology enthusiasts, then scan the code to add "Dynamic Points" to join the group chat!
Wonderful article is worth recommending!
One-click forwarding, poke and watch!