Exclusive interview with Bitmain: With an annual income of $2.5 billion from mining machines, why is it eyeing AI chips?

At the end of 2015, after a series of discussions, the company was about to take an important step: developing AI chips.

For Bitmain, this news is both unexpected and reasonable. Bitmain originally started out as a digital currency chip manufacturer, and now seeing the broad application scenarios of artificial intelligence, it is determined to attack the underlying architecture of artificial intelligence and enjoy the dividends of AI chips.

At the end of 2017, Bitmain officially announced its first AI chip, SOPHON BM1680, with the Chinese name "Suanfeng". In just one month, all related products were sold out. Bitmain completed the journey from zero to one in a year and a half, but its ambitions are not limited to this.

Strong Player

I don’t know when AI chips became a hot startup project. Bitmain is not the first company to discover the business opportunity, but it will definitely not be the last.

According to incomplete statistics, there are more than 40 companies worldwide that are making AI chips. Compared with the only few chip giants before, today's chip companies are developing secretly and violently.

"Let a hundred flowers bloom" is how Bitmain's product strategy director Tang Weiwei defines the current AI chip era.

Generally speaking, artificial intelligence can be divided into three layers: the basic layer, the algorithm (data) layer, and the application layer. "In the past, people valued algorithms and applications, but now they are beginning to explore the underlying things. More and more attention is paid to the role of chips." Tang Weiwei told Pintu Business Review (ID: pintu360) that the current data is large enough and the applications are rich enough to prompt talents to start thinking and find the most optimized infrastructure for industry applications. When general-purpose CPUs and GPUs cannot meet the needs of the artificial intelligence era, it is the time for AI chips to rise.

AI chips will also enter vertical sub-industries and become specialized chips for AI+ diversified industries. The fields that Bitmain is eyeing are security, Internet and urban big data.

If we take into account Bitmain's previous achievements in digital currency chips, the company has a very deep understanding of the chip industry. With its customized mining chips, Bitmain has become the world's largest supplier of Bitcoin chips and mining machine solutions, with revenue of $2.5 billion in 2017. Data shows that Bitmain has successfully developed and mass-produced a number of customized ASIC chips and complete systems, successfully mass-produced billions of chips in the field of digital currency chips, and has the most advanced 7nm process design experience. The most representative of these is its BM chip series used in encrypted digital currency mining machines.

It seems that Bitmain, which has already opened up the industrial chain in the chip industry, must be familiar with making AI chips, but AI chips are Bitmain's second venture, and it has not missed any pitfalls. Tang Weiwei told Pintu Business Review that AI chips and digital currency chips have a lot in common, such as the underlying physical design and process technology. But the upper system architecture needs to be re-studied.

Bitmain considered the choice of chip structure for quite a long time. If it used IP licenses from other companies, it would have limited choices in early 2016 and it would be difficult to find one that suited it. If it did not choose, it would have to make its own chip IP, which would inevitably be challenging and take longer. After careful consideration by the team, they finally decided to invest in research and development themselves. "Fortunately, the team was very determined in its decision-making process."

AI chip killer

BM1680 is a custom chip for tensor computing acceleration processing for deep learning applications. It is suitable for inference prediction and training of deep neural networks such as CNN, RNN, and DNN. Based on the BM1680 chip, Bitmain provides the Suanfeng SC1 and SC1+ board products and the intelligent video analysis server Suanfeng SS1, which is a new deep learning server. It is specially designed to provide powerful deep learning acceleration capabilities for a variety of application scenarios such as video surveillance and Internet image processing.

Tang Weiwei specifically mentioned that the chip has been optimized specifically for the security industry. For example, video decoding may not be available on traditional AI accelerators, but Suanfeng has this function.

The first and second generation products are relatively universal. Tang Weiwei said that generally speaking, the company's first generation products are trial products, so BM1680 is a large and comprehensive chip. The second generation chip will focus on security, and it is not ruled out that customized chips will be provided to customers in the future.

The reason why the security industry is so important is closely related to the scale of the industry itself. It is reported that the market size of China's security industry in 2017 was about 600 billion yuan, and the global market size reached one trillion yuan, which is huge; from the perspective of AI, security + AI has been mature and implemented, and it is most suitable for chips to re-transform it. In comparison, the autonomous driving scene has not yet been implemented, and the medical market is too large and too fragmented, which is not suitable for the rapid integration of AI for the time being.

In the security industry, AI chips are a topic worth discussing. With the deepening of the concept of AI+security, startups have become extremely popular in this field. Whether to make AI chips has become a question facing them.

In Tang Weiwei's opinion, most companies that are purely engaged in security will not get involved in AI chips, because the chip investment return cycle is too long and it is a high-threshold industry. " It takes several million US dollars just to invest in chips, which is a huge investment . Secondly, they need to attract talents to form a corresponding team. The most important point is whether they have the confidence to continue to invest for five years, because customers will not use your chip products only once."

In addition to this hard threshold, the high value of SOPHON chips also lies in the understanding of deep learning algorithms and chips, and there are barriers in both hardware and software. In addition, in terms of iteration cycle, SOPHON has set itself a goal of updating every 9 months, hoping to gain a leading advantage with the concept of "no speed is invincible".

Tang Weiwei told Pintu Business Review that the iteration cycle of traditional chip manufacturers is 18-24 months, and the performance of each generation only improves by 20-30%. Traditional chip development can no longer meet the needs of the industry, especially in China where Internet companies are developing rapidly.

Tang Weiwei believes that the current trend can be summarized as: software defines hardware. Almost all companies are moving towards services and data, products are transformed into data, and data generates value to further promote product growth. SOPHON uses a fast iteration method to map new data, algorithms and other models onto chips, using speed to grow together with customers. "In this case, I think the chip technology in the industry is likely to be subverted." Mature talents and technologies have become Bitmain's trump card in chips.

Industry Competition

Generally speaking, AI chips are divided into two categories: general-purpose and special-purpose. General-purpose chips include GPU (graphics processing unit) and FPGA (programmable gate array), and special-purpose chips are ASIC (application-specific integrated circuit). Both have different functions in different fields.

Strong ones such as Google's AI chip TPU, although using a dedicated chip ASIC, its purpose is to be universal. This is to allow Google to better promote cloud services and deep learning software library Tensorflow. In summary, everything serves the Google ecosystem. There are also pre-trained deep learning models in Google's cloud services to facilitate rapid deployment by people in different industries.

As the saying goes, only large companies build ecosystems, while small companies only focus on verticals. However, this statement may not be recognized by everyone. Tang Weiwei believes that it is very important for computing chips to build ecosystems, but computing chips are not all general-purpose chips. In vertical industries, it will take time to accumulate.

AI chips are also very popular in China. In the "2018 Artificial Intelligence Trend Outlook" report released by CBInsights, it is mentioned that in July 2017, the Chinese government stated that it would be equal to the United States in terms of intelligent chips by 2020 and become the world leader by 2030. CBInsights specifically mentioned the Chinese company Cambrian, saying that Cambrian promised to produce 1 billion processing units in the next three years and develop chips specifically for deep learning.

In Tang Weiwei's view, the AI ​​chip market is still a blue ocean. In most cases, companies are competing in a misaligned manner, and the market has not yet reached the era of real explosion.

The future of AI chips

With the advent of the artificial intelligence entrepreneurial boom, the AI ​​chip industry is also developing rapidly.

From the perspective of talent, the influx of more and more AI talents is a symbol of the industry's vitality. From another perspective, the cost of AI chip investment has been greatly reduced, and high-end chips in people's minds are gradually being accepted by entrepreneurs.

Tang Weiwei pointed out that the threshold for AI chip startups is indeed decreasing. But at the same time, the talent pool is increasing, the market is expanding rapidly, and industries such as AI, the Internet of Things, and blockchain all have a large demand for chips.

“If the time comes when everything is connected, everything will need a chip, and the market size of chips will be enormous.”

This scenario is not a fantasy, it will really happen one day in the future, giving rise to AI+ chips for various sub-industries. The era when only a few giants such as Intel and Nvidia dominated the industry is gone forever, and one or more giant companies may emerge in each field.

Tang Weiwei predicts that China's traditional enterprises have no experience in applying AI, but on the other hand, this also proves the huge potential of traditional industries + AI.

SOPHON BM1680 is a tensor computing chip and also an ASIC. Many people say that this AI chip is the Chinese version of TPU. Tang Weiwei himself disagrees with this statement: "We have never said that we want to replace anyone, but we show better cost performance and provide customers with more choices. No one wants to dominate this field, we are another choice."