In recent years, the Chinese government has vigorously developed the integrated circuit industry, and domestic CPUs have sprung up like mushrooms after rain. However, while some CPU companies promote their CPUs as being on par with Intel, such as Huawei claiming that the Kunpeng 920 outperforms industry benchmarks by 25% and has a better energy efficiency ratio than industry benchmarks by 30%, Intel still dominates the commercial market. So, how does the performance of domestic CPUs measure up, and where exactly is the gap with Intel? Here is a brief review of the following domestic CPUs by Tieliu.
X86 Camp:Lanqi, Haiguang, Zhaoxin
Since 2016, Lanqi Technology has collaborated with Intel and Tsinghua University to develop a secure and controllable CPU, and combined with Lanqi Technology’s secure memory module to launch a high-performance server platform, achieving the industry’s first hardware-level real-time security monitoring function, which can be used in the field of artificial intelligence.
Lanqi Technology’s official introduction:
The Jindai server platform is based on Intel’s Xeon general-purpose processor technology, integrating Tsinghua University’s reconfigurable computing processor (RCP) module and Lanqi Technology’s hybrid secure memory module (HSDIMM®), forming a dynamic security monitoring technology (DSC) to meet the urgent application needs of domestic data centers for high security and high-performance computing.
The Jindai CPU is essentially adding a so-called security module to the Intel CPU, making it CPU + ASIC, using Intel’s CPU directly, while the ASIC is developed by domestic units. In terms of performance, since it uses Intel’s X86 CPU, it can match Intel’s CPUs in performance. The issue lies in the fact that adding an ASIC will incur additional packaging costs, and the ASIC itself also has costs, so controlling costs is key. Additionally, in the case of using Intel’s X86 CPU with just one ASIC, whether it can guarantee 100% security and controllability is also a question that needs to be tested over time.
Haiguang is a joint venture established by AMD and Shuguang. Before the Ryzen was launched, AMD was commercially beaten down by Intel, hence it threw an olive branch to mainland China, establishing a joint venture and granting the joint venture access to the Zen source code and design documents. In terms of performance, Zen is at an international first-class level and is a competitive technology, which is reflected in the 800% increase in AMD‘s stock price over the past few years. It is said that Haiguang is already innovating its products, and if it can digest and absorb the technology while controlling costs, it could compete with Intel in the commercial market, which would be great.
Zhaoxin is a joint venture between Shanghai State-owned Assets and VIA from Taiwan. Due to being beaten in the market by Intel, VIA chose to form a joint venture in mainland China. During the 12th Five-Year Plan period, the joint venture undertook the National High-Technology Research and Development Program 01, receiving massive financial support. Zhaoxin has introduced VIA‘s technology, making several generations of improvements, with the latest KX6000 achieving a main frequency of 3.0G, reaching the performance level of the older generation AMD processors. If it can reduce costs, there is still potential in platforms like netbooks and thin clients. Admittedly, Zhaoxin’s CPU performance has improved significantly, but there is still a considerable gap in performance compared to Lanqi’s Jindai CPU and Haiguang’s CPU—Lanqi and Haiguang are estimated to score above 35 (SPECint2006, GCC), while Zhaoxin is estimated to be around 20 (SPECint2006, GCC, using ICC compiler can get close to 30, of course, Lanqi and Haiguang using ICC compiler will also see a spike in scores). For Zhaoxin, since it belongs to the X86 camp, the ecosystem is not a limiting factor; the performance and price of the CPU are the urgent issues to solve.
ARM Camp:ARM mini China、Huawei、Feiteng、Huaxintong
With the rise of chip fever in China, ARM also established a joint venture in mainland China.
Later, some netizens viewed the 51% stake of ARM mini China as a means for China to no longer fear foreign sanctions. However, ARM mini China 51% is essentially a shell company, and the establishment of this company is simply a way for Masayoshi Son to exploit the situation, and the idea that ARM mini China can allow domestic ARM camp manufacturers to be unafraid of sanctions is absurd.
Months ago, ARM decided to sanction Huawei, and the 51% state-owned ARM mini China aligned with ARM to jointly sanction Huawei, which sufficiently demonstrates that ARM mini China has no relationship with China’s realization of autonomous and controllable CPUs—it is merely a tool for collusion between domestic and foreign capital to exploit.
Years ago, ARM servers were hotly discussed in the industry, with the hype surrounding them comparable to the current RISC-V. ARM server CPUs were regarded favorably by many industry professionals, who believed they had the potential to replace X86 or erode part of the X86 server CPU market. Consequently, many well-known foreign companies such as AMD, Qualcomm, and Broadcom participated, with Huawei and Feiteng also joining this wave.
Huawei’s ARM server CPU has received support from national resources such as the National High-Technology Research and Development Program. Its ARM servers include Hi1610, Hi1612, Hi1616, Hi1620. Hi1616 is the Kunpeng 916, and Hi1620 is the Kunpeng 920. Huawei claims that the Kunpeng 920 outperforms industry benchmarks by 25% and has a better energy efficiency ratio than industry benchmarks by 30%.
In terms of chip design, Hi1610 and Hi1612 are integrated using A57, while Hi1616 is integrated using A72. Although Huawei officially claims Hi1620 is self-developed, based on the general rules of technological development, it should be a product built on the shoulders of giants—historically, Intel, IBM, AMD, ARM, and Apple have all undergone a process of iterative evolution in technology, where each generation of source code modification generally does not exceed 25%. The insertion of Apple is also based on modifications of the ARM public core—three generations of processors were developed based on ARM A8, A9, and A15. Consequently, ARM has stated that due to vulnerabilities in A8, A9, and A15, devices equipped with processors A4, A5, A5X, A6, and A6X also have potential risks. The notion of achieving everything solely through self-development is contrary to the general laws of development.
Feiteng’s ARM server development model is similar to Huawei’s. Feiteng has developed several chips including FT1500A, FT2000, FT2000plus, among which FT2000plus has a significant gap in single-core performance compared to Intel, but in multi-core performance, Feiteng targets Intel’s E5 processors. Currently, both Huawei and Feiteng have embarked on a path of digesting, absorbing, and innovating.
Huaxintong is a joint venture established by the Guizhou government and Qualcomm. In 2012, Qualcomm began planning to develop server chip business, initially intending to combine the energy-saving features of mobile chips with robust design capabilities to attract clients such as Microsoft and Google for their data centers.
Because Microsoft and Google aim to save operational costs and need to balance with Intel in price negotiations.
Afterward, Qualcomm poached talent from Intel, IBM, and AMD, establishing a design center in Raleigh, North Carolina, to kick off the ARM server chip program. At its peak, this team had over a thousand employees.
In January 2016, the Guizhou Provincial People’s Government and Qualcomm signed an agreement to jointly establish a joint venture, with Qualcomm providing its proprietary server chip technology to support the new company’s future development.
According to public information, Huaxintong’s initial registered capital is 1.85 billion yuan (approximately 280 million USD), with Guizhou holding 55% and Qualcomm holding 45%. Although this joint venture was once highly anticipated, Huaxintong announced its closure at the end of April 2019.
ARM server CPUs face three issues: performance, ecosystem, and price.
In terms of performance, although Huawei and Feiteng have used PPT to compare and even “beat” Intel, they are actually using 48-core and 64-core against Intel’s 14-core and 28-core. Domestic ARM CPUs still have a significant gap in single-core performance compared to Intel, thus can only compare multi-core performance by stacking core counts. However, the most critical aspect of a CPU is its single-core performance; practice has proven that strong single-core performance combined with hyper-threading is the optimal solution. Blindly stacking core counts without achieving single-core performance is a misguided approach.
In terms of the ecosystem, although it was widely rumored in the industry 4-5 years ago that the ARM ecosystem in servers would be rich, to this day, the promises made then remain largely unfulfilled. X86 maintains a remarkably solid advantage in PCs and servers.
Regarding pricing, chip prices are closely related to production volumes, and in this area, Intel has a clear advantage over domestic ARM CPUs. Furthermore, the common practice of stacking core counts in domestic ARM server CPUs further increases chip costs, making these domestic ARM server CPUs more expensive than Intel’s.
Qualcomm, AMD, and other foreign giants have abandoned ARM server CPUs, and Huaxintong’s closure stems from the fact that under pure commercial models, ARM server CPUs are not viable. Consequently, both Huawei and Feiteng are targeting the central procurement and secure controllable markets.
MIPS/SW64:Longxin, Shenwei
Longxin was established in 2001. The name Longxin might lead many to believe it is a CPU built with national effort; indeed, during the 10th Five-Year Plan, Longxin received significant national support. However, due to the “Hanxin incident” during the 11th Five-Year Plan, the scrutiny of chip projects increased, and national funding for chip projects became “very cautious.” It was Academician Li Guojie who used the computing institute’s funds to support Longxin’s research and development. After the 12th Five-Year Plan, the state resources strongly supported Zhaoxin and Huawei, while national projects like the National High-Technology Research and Development Program essentially cut off support for Longxin, leaving it to survive on its own through profits and investments from investors. Over the past 18 years, the state has only provided Longxin with over 500 million in funding. This financial strain has left Longxin lagging behind other domestic CPUs in manufacturing processes after being “weaned off,” such as the Longxin 3A2000 launched around 2015, which had a process of 40nm, while most domestic counterparts were at 28nm at that time. The Longxin 3A4000, released in 2019, was manufactured using 28nm, while domestic peers generally used 7/16nm.
Although financially constrained, in terms of the most critical single-core performance, Longxin has rapidly improved single-core performance to reach 20 points (SPECint2006) relying on its chip design capabilities, even while being 1-2 generations behind in manufacturing processes. Admittedly, Longxin has made significant progress in recent years, but the gap in CPU performance and ecosystem compared to Intel is still very clear; competing with Intel in the commercial market may still take 5 to 10 years.
SW was designed by relevant units and has been very successful in supercomputing, with the Sunway TaihuLight having ranked first in the TOP500 multiple times, and still ranking third now. After achieving success in supercomputing, SW began to focus on servers, with the SW3231 expected to be released in 2019 and the SW3232 in 2020. In terms of single-core performance, the SW3232 aims for around 25 points (SPECint2006), and if it can meet the expected performance and resolve the software ecosystem issues, then small-scale promotion is to be anticipated.
RISC-V
In recent years, RISC-V has been extremely popular. Some domestic organizations have chosen RISC-V in hopes of leveraging the current hype, believing that RISC-V has a bright future and may establish a certain ecosystem, allowing domestic units to be compatible with this existing ecosystem.
Currently, some voices in the media and industry believe that RISC-V has been overhyped, with its prospects overly elevated. This hype is very similar to the previous high-profile promotion of ARM’s entry into the server CPUs, proclaiming that ARM has limitless potential.
From a purely commercial perspective, playing X86 on smartphones and tablets, as well as playing ARM on PC and servers, is bound to fail.
Years ago, under Intel’s substantial subsidies, some domestic manufacturers collaborated with Intel to push X86 in mobile phones and tablets, resulting in Intel incurring losses exceeding 100 billion USD, leading to its drastic decision to withdraw.
ARM servers are retreating globally, with Qualcomm, ARM, and Broadcom demonstrating through practice that under a purely commercial model, ARM servers are not viable. In both domestic and international markets, X86 continues to dominate firmly, while ARM servers are only occasionally promoted by a few manufacturers using PPT to show their presence.
It can be said that in a purely commercial model relying entirely on the “invisible hand” for regulation, it is fundamentally impossible for newcomers to displace Intel and ARM from their stronghold markets.
Only through visible hand regulation can newcomers have a chance to thrive, gradually encroaching on the markets of Intel and ARM using strategies like “surrounding the city from the countryside” and “grinding a sword for twenty years.”
Therefore, for RISC-V, purely relying on market forces can only achieve some success in embedded applications, such as in the wave of the Internet of Things, but cannot make significant inroads in high-performance markets like PC and servers. In these areas, relying on market forces for RISC-V is as tragic as pushing X86 in mobile devices.
RISC-V does not address the issues faced by China in desktop and server CPUs; rather, it serves as a supplement in embedded applications. If RISC-V can rise as the media has hyped, it may dominate a portion of the embedded market, but it cannot resolve the issue of China’s high-performance CPUs being constrained by Intel.
Domestic CPUs Favorable Competition
Tieliu believes that the ideal state for domestic basic software and hardware is for all domestic CPU companies to design CPUs around a single instruction set; whichever CPU is better should be used, and software vendors should build ecosystems around this instruction set. However, since China has now gathered X86, ARM, Power, SPARC, MIPS, RISC-V, and many other instruction sets, it is unrealistic to forcibly merge teams through administrative directives. Therefore, a more realistic approach is to achieve survival of the fittest through market competition, similar to the path taken in the United States back then. For market competition to achieve survival of the fittest, a prerequisite is that there must be good rules; otherwise, only “bad money drives out good money” will occur.
First, it is not advisable to provide biased support and subsidies. In recent years, the state has increased its support for the integrated circuit industry, and some companies have received tens of billions in funding through projects like the National High-Technology Research and Development Program. While it is reasonable for the state to subsidize backward industries to help them develop, the methods of subsidy should also be prudent. Government funds can “assist for a while” and “send support,” but afterward, companies should be allowed to find their own way in the market. Especially for ARM servers, which have been proven commercially unviable by companies like Qualcomm, AMD, and Huaxintong, and have been shown to carry risks of sanctions in practice, it is inappropriate for national resources to be excessively subsidized, let alone to designate brands to be pushed into the secure market through administrative directives. Constantly relying on subsidies and administrative directives to “care for” will only allow companies to become flowers in a greenhouse, unable to grow into towering trees; they must be tempered in the market and grow strong after weathering storms.
Tieliu believes that applications are the ladder for the growth of domestic CPUs. The state should think creatively about how to help domestic CPUs be utilized, aiding their growth through this method rather than through the past approach of simply giving money. Even if subsidies are necessary, they should only cover the costs of wafer fabrication, and these subsidies must use local foundry processes, as this will help local foundries grow. CPU companies should not fabricate wafers using foreign foundries and then seek subsidies from the state.
Secondly, there should not be double standards regarding autonomous control. Admittedly, technology introduction is an effective means of mastering advanced technology, and there have been many successful examples in the past. For instance, after introducing the British Spey engine, the turbofan 9 was developed through digestion and absorption; after introducing CMF56, the turbofan 10 was developed; after introducing the Sukhoi 27, the J11 series and J15, J16 fighters were developed; and after introducing AP1000, the CAP1400 was developed. However, the key issue is that technology introduction generally involves a process of imitation (counterfeiting)—modifying the original design—then innovating. The ability to digest and absorb technology requires a certain level of capability and time. Immediately claiming autonomy after introducing technology is debatable. A certain unit has made a very insightful summary regarding the digestion and absorption of technology after introducing Russian fighters: “Modifying the original design must be done with caution.”
For technology introduction, the key lies in capability. Haiguang, Huawei, and Zhaoxin are all absorbing and digesting introduced technology, having acquired genuine materials. Haiguang has the complete CPU core and source code, Zhaoxin has obtained the source code from VIA, and Huawei has also received technology authorized by ARM. These three companies are all on the path of digesting, absorbing, and innovating on introduced technology, with the key being to form their own design capabilities and develop their independent products and IP. In contrast, Lanqi has merely packaged Intel’s X86 CPU with domestically developed ASIC modules without mastering core technology.
Here, I would like to elaborate further on X86 and ARM licensing. There are always uninformed individuals online who believe that X86 licensing has issues while ARM licensing does not. However, in reality, both X86 and ARM licensing have flaws and both have room for maneuver.
Regarding Haiguang, the background of the X86 cross-licensing between Intel and AMD is complex; cross-technology licensing complies with US laws and regulations. If Intel uses this to constrain AMD‘s business, it would face antitrust sanctions in the US, which would be even more severe and deadly for Intel, leading to netizens joking, “If the farming enterprise goes under, Intel will have to pay up.”
In today’s sensitive period, the US and China are also at a standoff, even though Huawei has been placed on the “entity list,” and even though ARM has announced compliance with Trump’s ban, there is still room for maneuver—Trump did not immediately implement a complete ban, but gave a 90-day grace period, which has not yet ended. After the G20 meeting, Trump has clearly eased his stance. It can be said that Trump’s goal is to “promote peace through war,” using the sanctions against Huawei as leverage to compel China to make concessions.
If the US and China really reach a breaking point, then all technology controlled by the US that China uses—software and hardware—will be at risk, and the entire Chinese commercial society may collapse. Even the Huawei Kirin, which many netizens revere, would be doomed.
Therefore, both X86 and ARM have flaws and room for maneuver. Some ARM supporters attack X86 licensing while ignoring the risks of ARM licensing, and pushing ARM chips into the secure market raises questions.
Finally, the secure market and the commercial market should progress together. First, it must be clarified that China is still unable to design, manufacture, and package domestic CPUs using domestic EDA, raw materials, equipment, and processes; therefore, autonomous control is a relative concept. This is also why Tieliu emphasized in previous articles that the pursuit of chip performance should not be excessive, but rather the pursuit of technological autonomy and the establishment of a “red industrial chain.” Regarding the secure market, it should serve as an incubator for China’s own IT technology system. History has proven that if China cannot achieve “whatever the foreigners have, we have,” it will be bullied. Establishing China’s own IT technology system is one of the material foundations for realizing the great rejuvenation of the Chinese nation. Tieliu believes that with successful examples already available, the pace of domestic substitution in special industries can be further accelerated.
In terms of the commercial market, the law of business should be allowed to function, and the government should act as a rule maintainer rather than excessively interfere in the commercial market. After all, the scale of China’s commercial market is enormous, and if subsidy methods are used to distort market laws, the required funds would be too high, making it unfeasible given the current unfavorable local debt situation. In the commercial market, if joint venture CPUs like Haiguang and Zhaoxin, which belong to the same X86 camp as Intel, can reclaim some market share from Intel, that would be very good. This would allow more profits to remain in the country and train local talent.
Tieliu’s view is that for the commercial market, as well as for special markets that require X86 ecology and performance, Haiguang is currently a choice with good performance and potential for technological absorption. The issue lies in cost control and whether the price is competitive. Regarding the establishment of a Chinese version of Wintel, Longxin and Shenwei can be anticipated, but in the short term, they lack competitiveness in the commercial market and will require a long-term battle of twenty years to succeed.
