Few people realize that digital currency mining, an emerging market worth nearly 20 billion yuan, is essentially a battle for chips.
"Financial" reporter Wu Yangying Hui / Wen Song Yu / editor
During the morning rush hour, a group of employees walked into the office building from the outside. They walked to the gate, straighten their neck and let their face appear on the small screen in the silver-gray pillar. The light is on, the gate is quickly opened, the whole process It took less than 1 second.
Here is Bitco's corporate headquarters outside Beijing's North Fifth Ring. Surprisingly, the world's largest digital currency mining company chose to display face recognition devices downstairs as a welcome ceremony – not in the tall iron frame. Black and gray mining machine arranged in densely packed.
'Our current position is 12 words, leading high performance computing chip company.' Bitcoin product strategy director Tang Weiwei told the Caijing reporter. Before joining BitChina, he was responsible for high performance computing and artificial intelligence in China. business.
The digital currency mining industry has formed a market worth nearly 20 billion yuan in the world. But few people know that this mysterious emerging industry is essentially a chip market. The essence of mining is the chip computing power developed by the competition mining company.
Due to the single application of the software, the mining machine chip has quickly passed the iterative evolution process in a short time - from CPU to GPU, FPGA, to ASIC (application specific integrated circuit) chip.
As the world's number one and second miner giants, Bitian and Jianan have applied mature ASIC chip technology to monopolize nearly 90% of the global market in the mining machine chip. This is the global chip industry, fresh Some market segments that are monopolized by Chinese companies.
China's chip is lagging behind the world's advanced level in the three major industrial chain of design, manufacturing, packaging and testing. Some areas even have a gap of 10 to 20 years with developed countries. This makes China's chip field over-reliant on import trade. Global research institutions IC Insights announced the global semiconductor top 20 in 2016, and no Chinese company was on the top 20.
However, this situation has changed in 2017. IC Insights announced the world's top ten chip design companies in 2017 (according to sales revenue), two domestic companies entered the top ten: Huawei Hisilicon and Ziguang Group respectively for 4.715 billion US dollars And the second and tenth place with $2,050 million in income.
Due to the accumulation of patents and insufficient software ecosystems, it is difficult for China to catch up with the leader in the traditional chip field that has fallen behind. In the chip field where there is not much historical burden and new technology breakthrough, China stands with other countries. On the same starting line, there is an opportunity to take the lead.
At this stage, the biggest hope of overtaking in China is in the field of artificial intelligence (AI) chips.
Google and Nvidia are the two giants in the field of artificial intelligence processing, but NVIDIA is mainly focused on GPU. The ASIC-specific chips in the field of artificial intelligence are still a blue ocean, and there are no giant companies that can monopolize the market.
In the AI field dedicated chips, in addition to Google, developers are mostly dynamic emerging companies, which are active in many Chinese companies, such as Cambrian, Horizon, etc. A technical director of a well-known Chinese chip company pointed out to the "Financial" reporter: ' The AI chip can be said to be the first in the US and the second in China.
At present, artificial intelligence chips are re-enacting the evolution path of mining machine chips – from CPU/GPU/FPGA to ASIC-specific chips. Digital currency mining companies also have the opportunity to expand the mature ASIC chip technology in the mining field to the field of artificial intelligence. On the list of China's top ten IC design companies released in 2017, Bitcoin jumped to second with annual sales of 14.3 billion yuan.
The transfer of artificial intelligence chips to ASICs will open up a new battlefield around the world. This is an opportunity for digital currency mining companies and an opportunity for Chinese chips.
The pain of the chip: Begging behind
China has a huge gap with developed countries on high-end general-purpose chips, and its dependence on foreign countries is very high.
China has suffered a painful blow from trade negotiations with the United States. US Secretary of Commerce Wilbur Ross announced on April 16 that in the next seven years, US companies will be banned from selling parts, goods, software and technology (mainly chips) to ZTE. The reason for the ban is that ZTE violated the US restrictions on the sale of US technology to Iran. When the ban came out, ZTE immediately fell into trouble.
The ZTE incident has stung the whole of China. The whole society is aware of the backward gaps and costs of the Chinese chip industry relative to the developed countries in the world. 'The chip involves many links, and the different fields are different. In some areas, China is difficult to go with the United States in 10 to 20 years. Competition. 'The technical director of a well-known Chinese chip design company told the Caijing reporter.
The chip industry is divided into three major industrial chain links: design, manufacturing, and packaging. 'These three links are said to be backward in China. In contrast, we may have the least behind the block testing, and production and design are lagging behind.' The technical director of the design company pointed out that these two links are precisely the highest threshold for chip technology.
In terms of design, China has a major shortcoming in high-end general-purpose chips, which is far from the developed countries. These chips mainly rely on imports, and their dependence on foreign countries is very high.
The global high-end general-purpose chip market is currently dominated by international giants. Intel chips monopolizes about 80% of the global market share on CPU chips, and AMD accounts for about 20%. NVIDIA monopolizes about 70% of the global GPU market share, Xilinx and Altera The two companies monopolize nearly 90% of the global market share of FPGA chips.
China's core chip self-sufficiency rate is extremely low. Everbright Securities Research Institute data shows that CPU, MPU in computer systems in China, FPGA/EPLD and DSP in general electronic systems, embedded MPU and DSP in communication equipment, and storage devices DisplayDriver in DRAM and NandFlash, display and video systems, the domestic chip share is almost zero.
In terms of chip manufacturing, the gap between China and the world is even more severe. In the three major aspects of chip design, manufacturing, and packaging, manufacturing is one of the weakest links in China. Caijing reported that Intel, Samsung and TSMC all announced It has already achieved mass production of 10nm chips and is ready to invest in 7nm and 5nm production lines. However, SMIC, the best manufacturer of chip technology in mainland China, is still attacking 14nm, and it is possible to achieve mass production next year.
China accounts for nearly one-third of global semiconductor sales, but its production capacity is only about one-tenth of the world. According to WSTS data of the World Semiconductor Trade Statistics Organization, global semiconductor sales in 2016 were US$338.9 billion, and China’s semiconductor sales were 1075. US$100 million, accounting for 31.7% of the global market. In 2016, the global IC manufacturing industry was US$112 billion, and China’s market share was only 11.6%. This data also includes foreign companies’ factories in China.
The backwardness of the two core aspects of chip design and manufacturing has made China have to rely heavily on chip import trade. China's chip imports are even as high as 1.57 times of oil imports, which is far ahead in industrial imports. China Customs Administration data shows that as of 2017 At the end of October, China's IC imports amounted to US$207.197 billion, while China's crude oil imports amounted to US$131.051 billion.
The Chinese government has realized the crisis of chip import dependence and tried to accelerate the development of domestic chips through the national power. In China Manufacturing 2025, the index of China's IC self-sufficiency rate is set to 40% in 2020 and 70% in 2025. .
At the same time, the National Integrated Circuit Industry Investment Fund was established. In the first phase of 2018, the funds raised by RMB 138.7 billion have been invested. Currently, the second phase of the fund is being raised, with a total amount of about 180 billion yuan to 200 billion yuan.
However, the injection of a large amount of funds cannot make China's chip industry go all the way. Technology accumulation and breakthroughs require more time. Global market research institute IC Insights pointed out that most of the technology used in the new chip production line in China's chip field is more than abroad. The main competitors are worse than two generations.
'The gap between us and foreign countries is not only because of the difference in design capabilities, but also historical reasons - such as patent accumulation, insufficient software ecological support. 'The above-mentioned technical director of China's well-known chip design company pointed out to the "Financial" reporter, 'There is not much history Bao Zheng, a field with new technological breakthroughs, China has a relatively good chance. '
Surprise outside the country: Chip battlefield monopolized by China
Few people realize that digital currency mining, an emerging market worth nearly 20 billion yuan, is essentially a battle for chips.
Bitcoin mining has existed for only five or six years. The industry has grown rapidly to a market valued at nearly 20 billion yuan and is expanding at a rate of 100%-200% per year. But few people realize that this 1. Block emerging markets are essentially a battle for chips.
Zhang Nanxuan and Wu Jihan, closely related to the birth of the world's first ASIC mining machines. They changed the direction of global bitcoin mining with ASIC mining machines and changed their own destiny.
When the two founders entered the Bitcoin world, they were just out-of-the-box students. At that time, the founder of Jianan Yuzhi, Zhang Nanyu, was a student at Beihang IC, and the founder of Bitland, Wu Jihan, just came from He graduated from Peking University for two years and worked as an investment bank risk analyst. In just a few years, they became the global leader in the global blockchain mining machine, and they have more than 80% of the world's mining machines.
The essence of Bitcoin mining is the computing power of the mining machine chip: Whoever develops the mining machine chip is more powerful, then using a single mining machine will dig more digital currency, thus occupying a larger global market.
This is a rare global chip market that has been monopolized by Chinese companies.
Bitcoin and Jianan Zhizhi are headquartered in Beijing and Hangzhou, China. The two Chinese companies firmly occupy nearly 90% of the global mining machine market. Interestingly, the rise of the Chinese mining machine is from the other side of the ocean. Provocative ignition.
In June 2012, the United States developed a 'Butterfly Lab' for bitcoin mining machines, claiming that they are preparing to develop an ASIC miner that is far superior to the current level. At the time, bitcoin mining commonly used GPU chips, while ASIC chips The computing power will be tens or even 100 times higher than traditional mining chips.
If the 'Butterfly Lab' develops an ASIC mining machine, it will have a computing power of more than 51% of the entire network - which means it will have the possibility of tampering with the bitcoin world.
In order to counter the possible monopoly of computing power, Zhang Nanyi, nicknamed 'Pumpkin Zhang', decided to develop ASIC mining machines and sell them all over the world. In January 2013, US Bitcoin core developer Jeff Gazik broadcasted a live webcast. Unpacking process: He opened a huge courier package from China, the world's first ASIC mining machine. This ASIC mining machine was developed by Zhang Nanzhen, who named it 'Avalon', King Arthur's legendary 'God The name of the island.
In July 2012, Jiang Xinyu, who graduated from the University of Science and Technology of China, nicknamed 'roasted cat', also announced plans to manufacture ASIC mining machines, and launched crowdfunding on bitcointalk in the Bitcoin forum, which issued 160,000 yuan at a price of 0.1 bitcoin. Shares, Wu Jihan bought tens of thousands of shares. In 2013, the ASIC mining machine developed by 'Baked Cat' was successfully launched, creating the first bitcoin mine in the world.
'Baked Cat' The company's 'Financial Report' released in June 2013 showed that its bitcoin mining revenue in the first quarter of 2013 was as high as 57,338.24 BTC. The virtual stock of 'Baked Cat' also rose rapidly, with the highest price per share. 0.1 bitcoin rose to 5 bitcoins. Wu Jihan received tens of millions of returns for this investment.
At this time, Wu Jihan was only one step away from starting his own mining company. He is not familiar with financial technology and bitcoin. He is a partner who knows the chip development technology. Fortunately, once on the streets of Beijing. The encounter led him to find the 38-year-old chip expert Jenke. The Jenke group was focusing on the set-top box business. On the second day of Wu Jihan’s second day, he studied the information of Bitcoin and blockchain in a few hours, and then decided Join.
Wu Jihan and Jenke Group jointly established the mining company 'Bitland' in 2013. The Jank Group developed a mining machine using ASIC chips in just half a year. In November 2013, the name was 'Antminer S1'. The mining machine is on the market.
As the company's ASIC chip miners continue to enter the market, the Bitcoin world has ushered in an 'arms race' that is madly rising.
Since the use of ASIC chips for power output in 2013, Bitcoin's average network computing power has maintained a high-speed trend. According to global consulting firm Frost & Sullivan, global bitcoin computing power has grown from 10.8Thash in January 2012 to At the end of 2017, 15 million Thash. In less than six years, the total network computing power has been turned over 1.38 million times, with a compound annual growth rate of 1335.7%.
In the fierce competition, time is life. Many mining machines that are being developed by mining machines are not even out of the factory, and they are facing the risk of being eliminated by the market. Bittin's first-generation mining machine 'Antminer S1' is extremely short-term research and development. Time is the key to its rise.
The S1 mining machine was launched in November 2013, just in time to catch up with the high point of bitcoin prices.
If it is a month or two later, Bitcoin may miss the opportunity to climb up and fall directly into the bottom. CoinMarketCap data shows that bitcoin prices peaked at the end of November 2013, reaching around $1,150. Then in a month's During the period of rapid decline, the lowest point in December fell to 560 US dollars, almost waist. From 2014 to 2015, Bitcoin entered the channel of shock down, the price is often in the low range of 200 US dollars - 300 US dollars.
In 2014, Bitcoin launched the mining machines of the ants S3 and S4. The S3 mining machine sold well, but due to the market downturn, the sales of the S4 mining machine was bleak, and the company's losses were not small. Fortunately, due to better cost control, Bitcoin can still support the development of a new generation of mining machines during the trough.
The 'roast cat' company, which started the production of ASIC mining machines, was in a fierce price change in this round of bitcoin, and the key stage of high-speed expansion of the whole network was left behind.
In October 2013, the second generation of chips for 'roasted cats' encountered difficulties and failed to develop in time.
Under the impact of a large number of competitors, the power of the 'roasted cat' mining machine dropped from more than 20% at the peak period to only 4%, and its virtual stock also fell below the issue price. After more than a year, 'roast cat ' Mysteriously disappeared on the way to the mine. No one knows where he is going from now on.
'Bear market is the opportunity to help Bitland to reach the first place.' A person close to Bit China told the Caijing reporter. In January 2015, Bittland launched a new generation of ant mining machine S5. Most of the opponents reduced their research and development due to the bear market. In this case, the ant mining machine S5 was successful in the market. This year, Bitian moved from a three-bedroom house to a new office, with more than 50 employees, and offices in Beijing and Shenzhen.
Only companies that have survived Bitcoin Winter will have the potential to grow rapidly in the next spring. Bitcoin and Jianan Zhizhi have seen opportunities to expand from mining chips to AI chips, entering AI from 2015 and 2016 respectively. The field of chip research and development.
'The mining machine business is a pillar of the company, but the income fluctuates with the rise and fall of the currency price, so it needs new business support. We have seen the wave of AI, and also the foundation of high-performance computing chips. 'Tang Weiwei, Director of Product Strategy, Bitcoin told "Financial" reporter.
Overtaking in a corner: new opportunities for artificial intelligence
The transformation of AI chips from CPU/GPU to ASIC-specific chips has created opportunities for Chinese digital currency mining companies to come to the fore.
The Jank group named the first-generation AI chip developed in Bitland as 'Sophon'. This inspiration comes from the popular science fiction "Three-body", which is a 'strong artificial intelligence' that the earth can't break through. .
Bitcoin's AI chip development path has a hidden similarity with 'Chikoko'. Bitcoin uses the ASIC chip technology used by Bitcoin mining machines in cross-border applications in artificial intelligence to develop AI deep learning chips based on ASIC technology. Just like a trio, the protons are unfolded in a low dimension and then folded in a high dimension.
Artificial intelligence has entered the stage of AI enabling industry, and demand is exploding. AI chips are in the stage of conversion from CPU/GPU/FPGA to ASIC-specific chips. Compared with the first three, the computing power of ASIC-specific chips Stronger, lower energy consumption.
According to Frost & Sullivan, a global consulting firm, CPUs are expected to lose market share in artificial intelligence training and reasoning applications, while ASIC chips will gain share of both markets.
ASIC chips will grow at a compound annual growth rate of 66% from 2017 to 2020, and the total market will grow from $1 billion to $4.6 billion, ranking first among all types of artificial intelligence chips.
Unlike the traditional chip market, which has been highly monopolized, there is no giant company that can't be challenged in the emerging market of ASIC chips. Currently only Google has a leading edge in ASIC chip technology. TPU is Google's dedicated integration designed for machine learning. Circuit, a class of ASIC chips.
Google pointed out in the official blog that the speed and power consumption of TPU chips are significantly better than traditional GPUs and CPU chips: 'TPU processing speed is 15 times to 30 times faster than current GPUs and CPUs; it is more energy efficient than traditional chip TPUs. The efficiency (TOPS/Watt) has increased by 30 times to 80 times.
China's digital currency mining company has accumulated many years in ASIC chips, giving them a first-mover advantage in this technological change of AI chips. Blockchain computing and artificial intelligence deep learning have similarities, relying on the underlying chips. Perform large-scale repetitive, parallel computing. Previously, digital currency mining companies' design experience accumulated in blockchain ASIC chip design is partially compatible with the technology required for neural network chip design.
Part of the experience of ASIC chips can be reused. However, the mining algorithm is not very complex, mainly because of its strong computing power. The AI algorithm needs universality and is a skill. ' Artificial intelligence company Rokid Beijing Lab Alab is responsible for Dr. Gao Peng told the reporter of Caijing.
Bitcoin believes that its R&D capabilities and experience in the field of digital currency chips can be replicated in the field of AI chips. It has achieved mass production of nearly 10 chip products from 2013 to 2018, with several billion productions, technical iterations. From 28nm to 16nm, with the latest 7nm design experience. At the same time, it has established tens of billions of calculations and 1 billion kilowatts of data center in the world. Currently distributed in China, Taiwan, Singapore, Europe, North America and other regions more than 2,000 worldwide 60% of employees are R&D personnel.
Bitcoin started research and development of the first generation of AI chip Sophon BM1680 in 2016. Tang Weiwei told Caijing that the Bit Continental AI team is hundreds of people, covering chips from algorithms to algorithms, software drivers, firmware and various Research and development of tools for deep learning.
AI chips and miner chips can be shared in many design processes and team capabilities, such as high-performance low-power technology and power management technology. But AI's algorithms are more diverse, algorithms, software, and tool chains need to be rebuilt. .
The Sophon BM1680 successfully streamed in April 2017 and took samples in the second quarter, which took nearly a year and a half. 'The first generation of chips was designed with a certain degree of redundancy, and the chip performance caused the physical design to be more complicated. So quickly go to correct, now Ability to quickly locate and update optimized designs. ' Tang Yuwei said.
In November 2017, Bitcoin officially launched the first generation AI chip Sophon BM1680. The first generation AI chip of Jianan Zhizhi is also in the process of research and development. Jianan Zhizhi’s prospectus shows that Jianan Zhizhi from 2016 The ASIC chip KPU, which began to develop artificial intelligence applications, is scheduled to be measured in the 2018 season.
For these digital currency mining companies, the challenge in the AI chip field has just begun. As an emerging intruder, as soon as they enter the battlefield, they must face several major problems in the chip field.
Counterattack puzzle: money, speed and technology
These companies have the power of rapid iteration and ample cash flow, but its ability to replicate into the AI chip field has become a core competence to be proved.
The long development cycle and excessive investment are two core problems that have long plagued chip companies. The market competition logic of digital currency mining and chip industry is not the same, which makes digital currency mining companies evolved with the chip industry in terms of capital and speed. Traditional different solutions.
'If there is a mistake in the direction of enterprise chip research and development, a large sum of money will be gone. ' Dr. Gao Peng, head of the artificial intelligence company Rokid Beijing Alab Lab, told the Caijing reporter.
One of the largest proportions of chip R&D expenses is the cost of filming. According to Jianan Zhizhi data, taking MASK tape as an example, 55nm is about 4 million yuan, 40nm is about 7 million yuan, and 28nm is about 1500 yuan. Around 10,000 yuan, the 16nm MASK is as high as 40 million yuan.
Different chips require different MASKs, and the chips in the development stage cannot guarantee the success of one-time streaming. A set of MASK needs to be modified repeatedly, the possibility of repeated investment, and the research and development costs have doubled.
One of the great advantages of digital currency mining companies is that the mining business brings sufficient cash flow, which allows them to spend more on chip research and development. The Bitian mainland co-founder Jank Group revealed to the media that Bitland’s annual revenue is as high as 2.5 billion. US dollar (about 16 billion yuan) or more. Jianan Zhizhi prospectus shows that the company's revenue in 2017 was 1.308 billion yuan.
'Our capital reserve can be invested on the chip and invested in the long-term. AI is a technical tool to change human productivity. It must not be a process of development after 3 years - 5 years. ' Tang Yuwei said.
The miner's chip generally undergoes a technical iteration for 9 to 10 months, which is much faster than the traditional chip. The chip industry has 'Moore's Law', which means that 18 to 24 months is a cycle, and the computing power is realized in each cycle. Some, or the number of transistors in the unit chip area doubled. The chip company's research and development iteration speed mostly follows this law.
Bit Continental AI chip research and development is only half of the traditional chip, the generation and generation interval is only 9 months. This rapid iteration ability, thanks to the fierce competition in the mining machine market in the past. Bitcoin network computing power in the past six years The compound annual growth rate of the year is as high as 1335.7%. It is equivalent to a single constant power plant. After one year, it can only dig up the equivalent of the original 7.5% digital currency production. This also makes the mining industry form 9 - 10 High-speed iteration efficiency of monthly update.
Taking the Itanium ASIC chip mining machine iterative process as an example, the iteration period is controlled from half a year to one year. Its 28nm ASIC chip mining machine iteration is faced with elimination after 2015-2016 sales, and the production cycle is over.
In its follow-up products, the first generation of 16nm is measured in the four seasons of 2016, the second generation of 16nm is measured in the four seasons of 2017, and the third generation of 16nm is expected to be mass-produced in the second quarter of 2018. The latest technology ASIC chip mining machine based on 7nm It is expected to be released in April 2018 and mass production in the third quarter.
Although the two digital currency mining companies have rapid iterative experience in the mining machine market, whether they can be completely copied into the AI chip field has become a core competence to be proved.
'Digging chips are faster in iteration, but different chip fields are different. Each segment has its own subdivision rules. 'The technical director of a well-known domestic chip design company pointed out to the Caijing reporter, 'From a technical point of view The mining chip has a single function, so it can be quickly iterated.
Jia Nan Zhi Zhi's AI chip KPU, target applications in smart home, smart city, intelligent monitoring and intelligent toy voice and image recognition functions and a variety of Internet of Things applications. According to Jia Nan Zhi Zhi prospectus, the current AI chip connection The total amount of pre-sale orders received by various customers is about 38.20 million yuan.
Bitland selects the point of entry for the artificial intelligence ASIC chip to provide an integrated solution from the underlying chip to the upper layer technology, algorithm and cloud computing. The preferred cut-in industry is security, and has already reached cooperation with public security systems and integrators in many provinces and cities in China. Internet + AI and urban big data are the focus of the next stage of development, digitizing a large number of people's livelihood social data, and finally using the AI mining to make people's livelihood information for urban decision-making.
'It is too early to comment on R&D capabilities. The artificial intelligence industry chain is divided into three blocks - computing power, algorithms, and data. The computing power was previously monopolized by foreign giants. At this stage, AI is the opportunity for us to overtake the corner. 'Bit Tang Weiwei, director of product strategy of the mainland, told the Caijing reporter.
Technological innovation will open up new tracks and create opportunities for latecomers to surpass their predecessors.