The most comprehensive review of domestic CPU performance: healthy competition and survival of the fittest

In recent years, the Chinese government has vigorously developed the integrated circuit industry, and domestic CPUs have grown like mushrooms after rain. How is the performance of domestic CPUs, and what is the gap with Intel? The author will make a relatively comprehensive and objective inventory of domestic CPUs.

X86 camp: Montage, Hygon, Zhaoxin

Since 2016, Lanqi Technology has cooperated with Intel and Tsinghua University to develop the Jintai CPU. The Jintai CPU is actually a so-called security module added to the Intel CPU, which is a CPU + ASIC. Since it uses Intel's X86 CPU, its performance is comparable to Intel's CPU. The problem is that adding an ASIC will lead to additional packaging costs, and the ASIC itself also has costs, so whether the cost can be controlled is the key. In addition, whether it can guarantee 100% security and controllability is also a problem that needs time to test.

Hygon is a joint venture between AMD and Sugon. Before the advent of Ryzen, AMD was suppressed by Intel in business, so it extended an olive branch to mainland China, established a joint venture, and granted technical information such as Zen source code and design documents to the joint venture. In terms of performance, Zen is world-class and competitive in the market, which can be reflected in the 800% increase in AMD's stock price in the past few years. It is said that Hygon is already carrying out iterative innovation of its products. If it can digest and absorb the technology and control the cost, it will be good to compete with Intel in the commercial market.

Zhaoxin is a joint venture between Shanghai state-owned assets and VIA of Taiwan, China. During the 12th Five-Year Plan period, the joint venture took over the 01 project of the National Key Basic Research Program and received massive financial support. Zhaoxin has introduced VIA's technology and made several generations of improvements. The latest KX6000 has reached the level of the old generation of AMD processors. If the cost can be reduced, it still has certain potential on platforms such as netbooks and thin clients. It is true that Zhaoxin's CPU performance has improved significantly, but there is still a big gap in performance compared with Lanqi Jintai CPU and Haiguang CPU.

ARM camp: ARMminiChina, Huawei, Feiteng, Huaxintong

As China is in the midst of a chip boom, ARM has also come to the mainland to open a joint venture. Some netizens see the Chinese side's 51% stake in ARMminiChina as a way for China to no longer fear foreign sanctions. A few months ago, ARM decided to sanction Huawei, and ARMminiChina was able to sanction Huawei in unison with ARM, which is enough to show that ARMminiChina has nothing to do with China's realization of independent control of CPUs.

Many years ago, ARM server CPUs were favored by many industry insiders and were believed to have the potential to replace X86 or erode part of the X86 server CPU market. As a result, a large number of well-known foreign companies such as AMD, Qualcomm, and Broadcom participated in the market, and Huawei and Feiteng also joined the wave.

Huawei's ARM server CPU has received support from national resources such as the National Key Research Institute of China's National Key Research Institute of Advanced Materials. Its ARM servers include Hi1610, Hi1612, Hi1616, and Hi1620. Hi1616 is Kunpeng 916, and Hi1620 is Kunpeng 920. Huawei claims that Kunpeng 920's performance exceeds the industry benchmark by 25%, and its energy efficiency is 30% better than the industry benchmark. As for the chip design itself, Hi1610 and Hi1612 use A57 for integration, while Hi1616 uses A72 for integration. Although Huawei officially claims that Hi1620 is independent, according to the general law of technological development, it should be a product that stands on the shoulders of giants.

Feiteng's ARM server development model is similar to Huawei's. Feiteng has developed several chips such as FT1500A, FT2000, and FT2000plus. Among them, although the FT2000plus has a large gap with Intel in single-core performance, Feiteng is comparable to Intel's E5 processor in multi-core performance.

Huaxintong is a joint venture between the Guizhou government and Qualcomm. Qualcomm will provide its proprietary server chip technology to the joint venture to support the future development of the new company. According to public information, Huaxintong's initial registered capital is 1.85 billion yuan, with Guizhou holding 55% and Qualcomm holding 45%. Although this joint venture was once highly regarded, Huaxintong announced its closure at the end of April 2019.

There are three problems with ARM server CPUs: performance, ecology, and price.

In terms of performance, although Huawei and Feiteng have used PPT to benchmark and even "beat" Intel, they are actually using 48 cores and 64 cores to compete with Intel's 14 cores and 28 cores. There is still a big gap between the single-core performance of domestic ARMCPU and Intel, so the multi-core performance can only be compared by the number of cores. However, the most critical thing about the CPU is the single-core performance. Practice has proved that strong single-core performance + hyperthreading is the optimal solution. If the single-core performance cannot be improved, blindly stacking the number of cores is a wrong way.

In terms of ecosystem, although the industry has been rumored for 4 or 5 years that ARM will have a rich ecosystem on servers, to this day, the boasts made back then are still just PPTs. X86's advantages in PCs and servers are extremely stable.

As for the price, the price of the chip is closely related to the output. In this respect, Intel has a clear advantage over domestic ARM CPUs. In addition, domestic ARM server CPUs generally use the method of piling up the number of cores, which further increases the cost of the chip. Therefore, these domestic ARM server CPUs are more expensive than Intel.

The reason why foreign giants such as Qualcomm and AMD gave up ARM server CPUs and why Huaxintong closed down is that ARM server CPUs cannot work under a pure business model. It is for this reason that Huawei and Feiteng are rushing to the central procurement and security control market.

MIPS/SW64: Loongson, Sunway

Loongson was founded in 2001. During the 10th Five-Year Plan period, Loongson received great support from the state.

The government has increased the intensity of chip project review, and the state has become "very cautious" in allocating funds for chip projects, which has delayed the arrival of state funds. It was Academician Li Guojie who used the funds of the Institute of Computing Technology to support the development of Loongson. After the "12th Five-Year Plan", Loongson can only make money on its own and rely on investors' investment to support itself. In the past 18 years, the state has only given Loongson more than 500 million yuan in funding. It is precisely because of the lack of funds that Loongson's chip manufacturing process generally lags behind other domestic CPUs of the same period after "weaning".

As for the most critical single-core performance of the CPU, Loongson relied on its own chip design capabilities and made rapid progress in single-core performance, reaching 20 points, even though its manufacturing process was 1-2 generations behind. Admittedly, Loongson has made rapid progress in recent years, but there is still a big gap between its CPU performance and ecology and Intel.

SW was designed by relevant units. After achieving success in supercomputing, SW began to focus on servers. SW3231 will be released in 2019, and SW3232 will be released in 2020. In terms of single-core performance, the target of SW3232 is about 25 points. If it can achieve the expected performance and solve the problem of software ecology, it can be expected to be promoted on a small scale within a certain range.

RISC-V

In recent years, public opinion and some people in the industry believe that RISC-V has been over-hyped. This hype is very similar to the high-profile hype a few years ago about ARM's entry into server CPUs and ARM's unlimited prospects.

It can be said that it is an impossible task to squeeze out the market share of Intel and ARM in their respective strong fields under the pure business model that completely relies on the "invisible hand" for regulation. Only through the regulation of the "visible hand" can latecomers have a chance to thrive.

Therefore, for RISC-V, relying solely on market forces, it can only make a difference in the embedded field, such as some embedded applications in the Internet of Things, but it cannot make a difference in the PC and server markets that require higher performance. Because relying on market forces to promote RISC-V in these areas is as tragic as promoting ARM in PCs and servers and promoting X86 in mobile terminals.

Domestic CPUs should have healthy competition

The most ideal state of domestic basic software and hardware is that all domestic CPU companies design CPUs around a set of instruction sets, and software manufacturers build an ecosystem around this set of instruction sets. A more realistic approach is to eliminate the weak through market competition, provided that there are good rules, otherwise it will only be "bad money driving out good money".

First, it is not appropriate to give partial support subsidies. In recent years, the state has provided great support to the integrated circuit industry. It is true that it is natural for the state to subsidize backward industries to help them develop, but subsidies should also be given in a proper way. Government funds can be used to "help" and "send" enterprises, and then enterprises should be allowed to make their own living in the market.

Application is the ladder for the growth of domestic CPUs. The country should think about applications and help domestic CPUs be used.

Secondly, there should be no double standards for self-control. Technology introduction is also an effective way to master advanced technology, and technology digestion and absorption requires a certain level of ability and time. For technology introduction, the real key lies in ability. Haiguang, Huawei, and Zhaoxin are all absorbing and digesting the introduced technology, and they all got the real thing. Haiguang has a complete CPU core and source code, Zhaoxin obtained the source code from VIA, and Huawei also obtained technology authorized by ARM. The key is to form your own design capabilities and develop your own independent products and IP.