What happened to China's supercomputers after the United States banned the sale of Xeon chips?

In April 2015, the U.S. Department of Commerce decided to ban the sale of Xeon PHI computing cards to China's four national supercomputer centers, which caused a huge uproar.

Some people swore that the Tianhe supercomputer would be finished because it would have no chips left.

Some people said that this sales ban is not feasible, and a few shell companies can solve China's demand for CPUs for supercomputers;

Some people also said that the US ban is a great opportunity for China to develop its own technology. Whatever the West blocks, China will achieve technological breakthroughs in this field in a few years.

Some people even said that China's supercomputers are just assembled products using foreign CPUs and lack technological content.

In August this year, the city of Leipzig in Germany released the 43rd Top 500 list of supercomputers in the world. From November 10 to November 12, Tianhe-2 still topped the list in the TOP100 released at the HPC China 2015 conference held in Wuxi.

Half a year later, 100P supercomputer construction plans such as Tianhe and Sugon have entered the public eye one after another, proving in practice that the US ban on Xeon PHI has not had much impact on China's supercomputer construction.

In fact, over the past half century, history has proved what Comrade Mao Zedong said: "Blockade it. Blockade it for ten or eight years, and China's problems will be solved."

Technological blockade cannot hinder China's technological progress

Since the 1960s, the United States and the Soviet Union have simultaneously imposed a technological blockade on China, forcing China to take an independent and self-reliant development path. Just as Khrushchev withdrew Soviet aid experts from China, forcing China to develop the "two bombs and one satellite" with its own capabilities, the simultaneous technological blockade by the United States and the Soviet Union forced China to develop an independent semiconductor industry with its own strength.

In the 1960s, Europe, the United States and the Soviet Union believed that China could not make transistors and could not produce its own transistor computers after the technology blockade against China. However, Kang Peng, a researcher at the 404th Teaching and Research Section of the 4th Department of the PLA Military Engineering Academy, successfully developed the "isolation-blocking oscillator" in 1962, which successfully solved the stability problem of transistors.

After the advent of the "isolation-blocking oscillator", Factory 109 began mass production of transistors, allowing China to enter the transistor era nearly eight years later than the United States, effectively countering the doubts of the West and the Soviet Union.

Then there were more good news:

In 1965, China's first independently developed integrated circuit was born in Shanghai, entering the integrated circuit era only five years later than the United States.

In 1972, the independently developed large-scale integrated circuit was born in Sichuan Yongchuan Semiconductor Research Institute, which was 4 years later than the United States in achieving the leap from small and medium-sized integrated circuits to large-scale integrated circuits.

By the end of the 1970s, China, which was an impoverished agricultural country with almost no heavy industry base, had caught up with the United States by relying on the dedication of its predecessors of "dedicating their youth and future generations" and the fighting spirit of "ten thousand years is too long, we must seize the day", and shortened the technological gap with the United States to 5 years.

Chinese scientific researchers and industrial workers have carried forward the spirit of self-reliance and self-improvement, built China's own semiconductor industry, and mastered the entire process technology from single crystal preparation, equipment manufacturing, to integrated circuit manufacturing.

At that time, only the United States and the Soviet Union mastered the entire process technology from single crystal preparation, equipment manufacturing, to integrated circuit manufacturing (Japan's technology was very strong at the time, but some areas were castrated by the United States).

After the reform and opening up, the wealth accumulated during the Mao Zedong era was squandered due to the pursuit of the theories of "market for technology" and "it is better to buy than to make, and it is better to rent than to buy". Technical talents either went abroad or went to teach at universities. Some were even transferred to guard the computer room. The technical talents cultivated during the Mao Zedong era were lost.

By the end of the 1980s, computer companies represented by Lenovo pursued the "trade, industry and technology" route and only engaged in the assembly of low-value-added microcomputers. It was already difficult for China to produce a computer with its own technology.

After China lost its ability to independently develop and manufacture computers, the US government strictly restricted the export of high-performance computers to China in 1989. In addition to the high purchase prices, the computers had to be placed in a transparent glass room and monitored by Americans. The keys were in the hands of Americans, and each use had to be approved by the Americans.

After learning from its mistakes, China restarted its high-performance computer research and development project, and Galaxy 2 was born in 1992.

In 1993, Dawn 1 was born. Three days after the successful development of Dawn 1, the West lifted the export restrictions on high-performance computers to China.

Therefore, technological blockade is actually a good thing. The history since the founding of the People's Republic of China can illustrate this: as long as the Western technological blockade exists, domestic compradors and foreign forces will not be able to kill independent technology by means of " market for technology " and " it is better to buy than to make , and it is better to rent than to buy ".

Then, China's human, material and financial resources will all be invested in the research and development of independent technologies. With the support of China's most complete industrial sectors and solid industrial foundation in the world, as well as China's huge market and abundant resources (total resources, not per capita resources), it will be possible to "block whatever is blocked and solve whatever is solved." Therefore, in the long run, technological blockade is a good thing.

What is the impact of the ban on sales on China's supercomputers?

Just after the United States banned the sale of Xeon PHI for supercomputing to China, the National University of Defense Technology successively released the independently developed "Mars" and Matrix 2000.

For a detailed introduction to "Mars", please refer to the article by a computer engineer from the Chinese Academy of Sciences: "Why is Mars, a domestic chip that wants to challenge Intel, so questioned?"

Mars will be taped out before the end of this year. Mars is compatible with ARM V8 instructions, and its microstructure is named Xiaomi, which is a microstructure superior to ARM cortex A72.

Mars has 64 cores, a main frequency of 2G, and a theoretical double-precision floating-point computing capacity of 512Gflops;

In order to cope with memory-intensive operations, Mars has 32M L2 cache, 128M L3 cache and 16-channel memory, 205G/s theoretical memory bandwidth, and 32 PCI-e 3.0 channels on the chip, making it very suitable for serving as the main core of supercomputing nodes.

It is planned to be manufactured by SMIC, with a process technology of 28nm and a power consumption of 120W. It is not easy to achieve a power consumption lower than that of E5 when the process technology is two generations behind Intel.

Although the single-core performance of "Mars" is relatively weak, relying on the advantage of the number of cores, in the Spec 2006 benchmark, the multi-core integer score reached 672 and the floating point score reached 585, which is enough to compete with Intel's Xeon E5 2699v3.

Mars is clearly designed to replace the E5 in compute nodes.

Matrix 2000 uses a 40nm process technology, has 16 cores, a main frequency of 1G, a double-precision floating point of 2.4Tflops, and a power consumption of 200W. The performance-power ratio of Matrix 2000 not only significantly surpasses the first-generation Xeon PHI (2.4T/200W vs. 1T/300W), but is also slightly better than the second-generation Xeon PHI (2.4T/200W vs. 3T/300W). Moreover, this is the result achieved when Matrix 2000 uses a 40nm process technology, which is far behind the 14nm process technology of the second-generation Xeon PHI computing card.

Therefore, Matrix 2000 can completely replace the Xeon PHI computing card, and there is still great potential to be tapped.

According to the information released by the National University of Defense Technology, because the main I/O structure of Tianhe-2 is retained, the main processor of the computing node of Tianhe-2A still uses E5-2692 V2 (using "Mars" requires a redesign, and the change is too big, so it is safer to proceed in steps). The number of computing nodes has increased to 18,000. According to the calculation that one computing node requires 2 E5s and 3 accelerators, Tianhe-2A needs 36,000 E5s and 54,000 Matrix 2000s. The theoretical floating-point peak of 54,000 Matrix 2000s alone can reach 129.6Pflops.

（GPDSP）

It is obvious that as long as the National University of Defense Technology can properly ensure the efficiency of GPSDP, GPDSP will be very competitive. After replacing the Xeon PHI computing card with Matrix 2000, the overall efficiency of Tianhe-2A will be greatly improved compared with Tianhe-2, and the theoretical floating-point performance will get a qualitative leap, and it is expected to once again retain the top spot in the TOP500 ranking.

The new generation supercomputer system designed by Sugon - Sugon 7000 will fully adopt independent technologies from processors, high-speed communication networks, large-scale storage systems, system software to application software. It will be safe and controllable, and its performance will exceed 100P.

A supercomputer fully adopts independent technology, and its performance parameters such as performance-to-power ratio, overall efficiency, and theoretical peak value are unparalleled - the peak value of the supercomputer system is almost twice that of Tianhe;

The overall efficiency of the supercomputer system is 16% higher than that of Tianhe-2;

The performance-to-power ratio of the supercomputer system is 2.5 times that of Tianhe-2. According to the Green500 ranking released in June 2015, this supercomputer can be directly ranked in the top five.

After the supercomputer is completed, it will provide computing and technical support services in the fields of medical services, biomedicine, marine science, modern agriculture, oil and gas exploration, climate meteorology, financial analysis, information security, industrial design, animation rendering, etc., and provide a platform support for China's scientific and technological innovation and economic development.

Therefore, after the United States banned the sale of Xeon chips, it only caused a delay in the upgrade of Tianhe-2, and did not have much impact on the development of other supercomputers in China. Moreover, after the ban, it strengthened China's confidence in developing independent technology, which is more beneficial to the development of independent technology.

Banning Xeon chips from China could backfire

As early as 2013, the National University of Defense Technology applied for a number of GPDSP patents, which may be used as a substitute for Xeon PHI.

Therefore, it is impossible for the US intelligence agencies not to know that the National University of Defense Technology is developing GPDSP to replace Xeon PHI. So why is the US still doing this useless work?

Because the ban on the sale of Xeon chips is not aimed at China's supercomputer construction, but is more political and involves intelligence gathering.

Because the ban on the sale of Xeon chips is not aimed at China's supercomputer construction, but is more political and involves intelligence gathering .

The ban on the sale of Xeon chips is part of the United States' political combination of hitting China. As China's comprehensive national strength continues to improve, especially since China established an air defense identification zone in the East China Sea, reclaimed islands in the South China Sea and deployed increasingly powerful naval and air forces, it has become a major concern for the United States' Asia-Pacific rebalancing strategy.

The United States has used various restrictions, including economic, public opinion, and technological, to warn China not to go too far, using an invisible stick to hit China, and the ban on the sale of Xeon chips is obviously part of it.

But banning the sale of Xeon chips is not a good move - after losing China as a major customer, not only did Xeon chip sales drop by a factor of two and costs increased, it also caused huge economic losses to Intel.

More importantly, the sales ban has made some people in China who believe that "it is better to buy than to make" feel devastated. They can no longer hold back on independent technology as they did in the past. This has strengthened China's determination to replace foreign technology with independent technology, and to concentrate more resources and energy on replacing Xeon PHI computing cards with Matrix 2000 and domestic many-core chips, thus accelerating the process of localization of Chinese supercomputers in terms of chips.