Bitmain's Zhan Ketuan: In the next 3-5 years, the price of single-channel video structure will drop by 100 times, which is a reasonable level
Latest update time：2019-10-28
Reads：
▲Click above Leifeng.com Follow
In the next 3-5 years, the market size of video structuring will be 10,000 times that of today, and a 100-fold drop in the price of single-channel video structuring is a reasonable level.
Text | Bao Yonggang
According to Leifeng.com, one day before the 2019 Shenzhen Security Expo, Bitmain released its third-generation intelligent server SA5 in Shenzhen. This AI server is equipped with the BM1684 AI chip that was just released in September and is mainly aimed at the field of video and image intelligent analysis. Bitmain Chairman Zhan Ketuan said at the press conference that the market size of video structuring will be 10,000 times that of today in the next 3-5 years, and a 100-fold drop in the price of single-channel video structuring is a reasonable level.
The video structured market size will increase 10,000 times in 3-5 years
The reason why video structuring is needed is that the original video images collected by the camera are unstructured data and cannot be directly read and recognized by the computer. If you want to make video images play a better role in smart cities and safe cities, you need to use intelligent video analysis technology to structure the video images. Taking the security field as an example, it is necessary to perform intelligent analysis on the original video to obtain three types of key information. First, identify the moving target, then further determine whether the moving target is a person, a car or an animal, and their characteristics, and finally analyze the movement trajectory of the identified target.
This involves processing methods such as time-space segmentation, feature extraction, object recognition, and deep learning. This process is very complex and requires strong computing power support, so video structuring has not yet been widely used.
Zhan Ketuan said that today, video structuring at the level of thousands of channels is only sporadic in the market. One of the video structuring projects we won only had a few thousand channels, which is already a very large scale.
But in the future, all camera data will be structured and structured in the cloud.
He also pointed out that the number of security cameras shipped worldwide each year is about 100 million. Based on the five-year inventory, it is 500 million, and most of them are in China. Therefore, the scale of the video structured market will be 10,000 times that of today in the next 3-5 years, but it will not be able to be sold at today's market price. Therefore,
from the perspective of technology and products, the price of single-channel video structured in the market will drop in the next 3-5 years. I think a drop of 100 times is a reasonable level.
Wang Chenguang, President of Bitmain’s AI Business Line
Improve performance by 8 times and reduce computing cost
Since the cost of video structuring needs to be significantly reduced, the reduction of the underlying computing power cost is crucial. One of the important reasons for the launch of SA5 is to meet the needs of video and image structuring. Wang Chenguang, President of Bitmain's AI Business Line, said at the press conference, "Bitmain's SA5 intelligent server is mainly aimed at the field of video and image intelligent analysis. It can be equipped with third-party algorithms to run video structuring, face detection and recognition, vehicle detection and recognition, object detection and recognition, video detection and other types of security and Internet video intelligent analysis applications. It is currently the most mature and stable intelligent server in the field of artificial intelligence."
It is understood that the BM1684-based SA5 is a 19-inch standard 2U server with three built-in intelligent analysis units, including 24 BM1684 chips, which can provide a peak computing power of 2.2*24TFlops for 32-bit floating point and 17.6*24Tops for 8-bit integer. In terms of power consumption, the overall power consumption of SA5 is less than 700W, and it can provide up to 422T computing power in INT8 mode, which is about 8 times higher than the previous generation of products.
At the press conference, Bitmain demonstrated the effects of 384-channel real-time video streams based on CPU and GPU. It can be seen that GPU has more advantages in video processing, but TPU can bring better performance.
Wang Chenguang said that a single SA5 has the processing capacity of 768 channels of 1080P video at 30 frames per second or the detection and analysis of 12,000 face images per second, which can greatly improve the efficiency of video/image intelligent analysis in industries such as smart cities, smart security, and the Internet.
In addition, SA5 adopts a pluggable modular design, with built-in core modules such as the main control unit (MPU), computing power unit (AIU), and storage unit, which can facilitate later operation and maintenance. The main control unit is an intelligent management and scheduling system based on the X86 architecture, which can realize algorithm deployment, system upgrades, computing power performance monitoring, energy consumption management, and dynamic load balancing of computing power. It can provide management and data exchange for the AIU unit, and a 20G network interface to meet the input and output of computing power data.
The computing unit is the core of the server. SA5 integrates BM1684. A single unit integrates up to 8 BM1684 chips. The whole machine can be configured with multiple units as needed. The storage unit of SA5 is an optional unit. If the computing unit is not fully equipped, a storage unit can be added.
As for BM1684, it is positioned to focus on artificial intelligence reasoning for cloud and edge applications. It uses TSMC's 12nm process. Under the premise of typical power consumption of only 16 watts, the FP32 precision computing power reaches 2.2TFlops, and the INT8 computing power can be as high as 17.6Tops. Under Winograd convolution acceleration, the INT8 computing power is further increased to 35.2Tops. At the same time, the BM1684 chip also has a variety of encryption and decryption algorithms built in, which can protect the security of customers' algorithms and models, and create a safe and reliable computing environment.
Zhan Ketuan revealed that the fourth-generation BM1686 AI chip based on 7nm process will also be released next year.
Algorithm companies are all friends
While reducing the cost of computing power, only better integration with algorithms can bring about efficiency improvement and ultimately achieve a 100-fold reduction in cost.
However, Zhan Ketuan emphasized that Bitmain focuses on providing computing power chips and computing power hardware, and does not sell algorithms on the market. All algorithm manufacturers are our friends, our ecological partners, and our customers.
Therefore, in order to enable developers and users to quickly and easily implement intelligent applications such as video structured analysis and face recognition, Bitmain provides a one-stop toolkit including the underlying driver environment, compiler, inference deployment tools, etc. - BMNNSDK (BITMAIN Neural Network SDK).
It is reported that through BMNNSDK, developers can automatically compile and convert models developed and trained based on deep learning frameworks into applications that can run efficiently on the SA5 high-performance computing platform. Currently supported deep learning frameworks include TensorFlow, Caffe, mxnet, ONNX, Caffe2, etc., and support neural network models such as CNN/RNN/DNN.
Unlike Bitmain, many AI chip companies provide integrated hardware and software solutions in order to achieve better hardware and software integration.
Why did Bitmain choose not to sell algorithms but to cooperate with algorithm companies?
In an interview with Leifeng.com after the launch of BM1684, Zhan Ketuan said that the integration of hardware and software in the field of AI may bring cost and performance advantages in terminal cameras, but the cloud is a general high-performance computing center platform that must be able to support a variety of different algorithms.
Bitmain has also been thinking about this for a long time. In the end, we decided to do what we should do and what we should not do, providing chips and computing hardware, and cooperating with more algorithm manufacturers to build an ecosystem.
At the SA5 launch event, Bitmain’s partners CloudWalk, Isaac, Shenmi, and Qianshitong shared how they collaborated with Bitmain. All four speakers mentioned that the toolchain and services provided by Bitmain allow them to use the computing power of hardware conveniently and efficiently, but they also said that they can improve the accuracy of the algorithm, but to achieve sufficiently low costs, they also need Bitmain to provide lower-cost hardware.
Since the cooperation mode has been chosen, the construction of the ecosystem is naturally more important. At the press conference, Bitmain signed contracts with 17 ecological partners including CloudWalk, Isaac, Shenmi, Qianshitong, HiSilicon, SmartEye Technology, Huafu, Nanze Intelligence, Huazun Technology, Landian Technology, Lingxiang Technology, Hangkang Information, Weishiqing, Warehouse 37, Kaus, Xinhe Huitong, and Niu Intelligence to jointly build an open ecosystem.
Leifeng.com Summary
The implementation of AI technology requires the support of higher computing power, which is also an important reason why AI chips have received so much attention in recent years. One
of the important reasons why Bitmain has achieved good results in the field of AI chips is that it has been focusing on the design of dedicated chips since its inception. Since the first generation of AI chips were taped out in 2016, it has continued to accumulate experience in the field of AI chips. As for
whether the approach of only providing hardware and cooperating with a large number of algorithm companies can be successful, it remains to be seen. However, at present, hardware based on Bitmain's computing power has made good progress in the fields of city brain, smart security, smart communities, etc.
Previous recommendations
▎How did Pinduoduo’s market value surpass JD.com?
Are you still watching?
京公网安备 11010802033920号