Bitmain: AI chip updated every 9 months, Sophon 2nd generation is coming soon

Bitmain, which has always been mysterious and low-key, gladly accepted an interview at GTIC (Global Technology Innovation Conference).

This was an interview that was upgraded from an exclusive interview to a group interview with five companies. The topic was AI chips, which is a topic that all media outlets feel has a technical threshold to talk about, but one that must be crossed due to its popularity.

The chip industry has already shown a trend of many companies standing on the same footing, but the launch of AI chips by companies such as Huawei HiSilicon and Cambrian still boosted the enthusiasm of the industry, capital and media.

Bitmain proposed that AI chips should be updated every nine months. "Compared with the iteration speed of traditional chips, AI algorithms are iterated faster. We target the needs of the latest algorithms and the common foundations of neural network algorithms and quickly put them on the chip," said Tang Weiwei, director of product strategy at Bitmain.

In 2015, Bitmain was quietly preparing AI chips. They used the most cost-effective 28nm technology at the time. It is certain that Google was doing the same thing at the time, because at the Google I/O Developer Conference in May 2016, Google officially announced the existence of TPU for the first time.

Compared with general-purpose CPUs and GPUs, TPU is a fast and energy-efficient ASIC. ASIC is an integrated circuit designed for specific applications. Tang Weiwei compared digital currency chips and AI chips when answering questions from Yicai Technology. According to him, even though they are both ASICs, the logic of the two chips is different. The design of virtual currency is based on hash algorithms, and the design of AI chips is related to neural networks, but the underlying physical design has something in common.

Bitmain's first AI chip, named Sophon BM1680, was launched in November 2017. It was taped out in April 2017 and samples were available in the second quarter of that year.

In Liu Cixin's "The Three-Body Problem", Sophon is a super-intelligent machine from an alien planet that can change the fate of mankind. Sophon can develop a very strong learning ability through two-dimensional particles. It can be seen that this chip is also expected to use artificial intelligence to change all walks of life.

This chip has now been put into mass production, taking into account both training and reasoning functions, but mainly inference. He believes that training and reasoning should be two different platforms, and in the future, the focus will still be on reasoning. "High-performance computing involves many fields, so we decided to enter the field of AI at the end of 2015, especially the field of deep learning. On the basis of some existing high-performance computing chips, hardware, and some software algorithms, we have also introduced a large number of professional talents in AI." Tang Weiwei said.

According to the prediction of the third party Tractica, deep learning on GPU is indeed the current situation in the industry. With the maturity of deep learning algorithms and the rapid evolution of system architecture, the application of ASIC will exceed GPU by 2020. In fact, in addition to GPU manufacturers, FPGA (field programmable gate array, which can achieve semi-customization) manufacturers are also trying AI chips.

Algorithm applications change so quickly. Can dedicated chips keep up with the speed of algorithm changes? How to compete with existing chip giants? Tang Weiwei said that in fact, FPGA will be used for design before testing tape-out, but it will not be launched as a mass-produced product. He also believes that there will be more cooperation than competition in the future.

"AI algorithms and applications are not fully mature yet, and there are countless new iterations of AI algorithms. If AI algorithms are already iterating rapidly, chip companies cannot be so slow. So we hope to be a very fast chip company, but also a very stable and fast company," said Tang Weiwei.

The second quarter of this year is when Bitmain Sophon 2 is expected to be launched.

Bitmain Sophon official website: https://www.sophon.ai/