​BITMAIN launches SOPHON Tensor Processors and AI Solutions

On November 8, 2017, AI World 2017 (http://www.aiworld2017.com/en/), titled “AI Power Universal Leverage” was held in Beijing. Bitmain was honored to be invited to the event, and its cofounder and CEO, Micree Zhan, gave a keynote speech titled “Planet-scale Computing Driven by AI,” introducing the bright future of the AI revolution, big data, and massive planet-scale computing. He also officially unveiled the world’s first tensor computing ASIC, Sophon BM1680, proudly announcing that it’s already in mass production. At the event, Bitmain showcased Sophon BM1680 chips, the Sophon SC1 card, and the Sophon SS1 server. They also presented a live demo of intelligent video analysis run on the Sophon SS1 system. The official website of Sophon, sophon.ai, went live today.

Company Background and Vision

In the past few years, with the rapid growth of data generation and break-throughs in deep learning algorithms and computer architecture, AI has set off a new wave in the global IT industry, causing a fundamental shift. Not only does this create new opportunities; it also brings new challenges. To meet AI‘s increasing demand for computing power, current computer arcitecture faces challenges in the areas of power consumption, performance, memory bandwidth, the slow growth implicit in Moore’s law, and more. Computer architecture desperately needs innovation.

“AI drives the new generation of planet-scale computing and computer architecture innovation. In terms of images, voice recognition, and natural language processing, AI applications are going ever deeper and continue to drive innovation in the industry,” said Micree Zhan, CEO of Bitmain.

“As the global leader of digital currency ASICs and solutions, as well as the world’s largest provider of Bitcoin mining ASICs, we strongly believe that AI will soon have applications in every industry, making the whole world more intelligent.”

Bitmain is committed to providing the most cost effective and least power consuming AI computing power based on its strong ASIC design experience, agility, and complete system design capability. We will focus on fully customized design for AI applications and optimized hardware and system level solutions in order to lower the barrier to the realization of AI+ and the dream of seeing AI implemented everywhere.

Consider the example of the Sophon SS1 server solution we announced today. It is a complete integrated system with many AI algorithms. Customers can simply plug it in and launch the video analytics demo; furthermore, the applications can be easily integrated into large video analytics systems by calling algorithm APIs. Customers may also choose the Sophon SC1 accelerator card based on the Sophon BM1680 tensor processing unit.



Sophon BM1680, Sophon SC1 and SC1+ Cards

Sophon BM1680 is a customized tensor computing ASIC for AI applications and deep-learning training and inference, supporting CNN, RNN, and other deep neural networks.

The BM1680 chip is made of 64 NPUs. Each NPU has 64 EU operation units. NPU adopts improved systolic architecture and a DDR4 interface.


Based on Sophon BM1680, Bitmain provides Sophon SC1 and SC1+ cards. The SC1 card has one high performance BM1680 chip and the SC1+ card has dual BM1680 chips which are connected through the high speed SerDes Chiplink, introducing a new level of acceleration for deep-learning.




Software Interface and Full SDK Pack for Sophon Chip and Card Solutions


Sophon SS1 Server System

Sophon SS1, a specific server system for intelligent video and image analytics, is powered by BM1680 and SC1, aiming to provide a holistic solution for multiple AI application scenarios.

Sophon Video Analytics AI Demo

This demo shows actual production scale video analytics applications for the surveillance industry. It showcases motor/non-motor vehicle and pedestrian detections; providing rapidly iterated structural video analytics APIs.


The official English and Chinese versions of the sophon.ai website are live. We welcome customers to visit us to learn more about our latest products and solutions. All Sophon series products are currently available for order.

SC1/SC1+ Features and Price:




1 x BM1680

2 x BM1680


2TFlops FP32

4TFlops FP32




Power consumption



Official price

¥4620 / $589

Not available for sale

SS1 Featurs and Price:


Intel E3 1275V6, 4 Cores, 3.8GHz (Max Turbo 4.2GHz)

DL accelerator

2 x SC1+

Official price

¥30400 / $3869

Bitmain was founded in 2013 and had strong R&D capability in HPC field. It had successfully mass production many ASICs and systems with the extensive manufacturing experience of advanced IC processing technology, 28nm, 16nm, and 12nm. The most representative case is BM series chips for digital currencies mining. According to the official data, the 5th generation of Bitmain’s mining chip, BM1387, is the world’s least power consuming and highest performance accelerated computing chip. It can perform 80 billions of operations per second and has been shipped with the quantity of over a couple billions.

Riding on its success in cryptocurrrency, Bitmain has already made the plan to enter the more and more popular AI market. Accompanying the development of AI algorithms, AI applications, and the ecosystem of ASIC industry, the importance of AI ASIC chips will become more and more important.

In the future, Bitmain will continue to develop AI, and expand the investment in AI technology and talent in order to speed up the AI development. According to Micree Zhan, CEO of Bitmain, entering AI market is an important step for Bitmain to build its global brand. Starting from developing ASICs for digital currencies, manufacturing miners, and setting up the mining pools, Bitmain succeeded without any external support. Bitmain built the complete supply chain and continued to increase its unique relevance in the industry. And now, it is the best timing for Bitmain to enter AI market. Based on the solid foundation of the first generation product, looking toward the future, Sophon would continue to enhance the hardware technology and invest deeply in AI. We are planning to launch the second and the third generation products with much higher performance per watt using the most advanced 12nm FinFET technology to expedite AI revolutions.