Chinese     /     English


Copyright:Jiangyin NPD Machinery Co., Ltd.     苏ICP备18054409号 

AGM Micro has dabbled in AI terminal chips

Industry news
2019/01/15 12:17
Page view
Recently, in the field of AI chip "cold" adds a new player, AGM ( is launching a MCU based on ASIC is force to accelerate chip, cost-effective embedded MCU market, used for iot AI re
Recently, in the field of AI chip "cold" adds a new player, AGM ( is launching a MCU based on ASIC is force to accelerate chip, cost-effective embedded MCU market, used for iot AI reasoning work force to accelerate the edges.
In recent years, deep neural network (D**) has achieved unprecedented success in image recognition, natural language processing and other aspects, and promoted the rapid development of relevant industries. However, the deep neural network used in these applications has a large number of parameters, and both model training (trai* I *g) and inference (I * crawled *ce) require a lot of calculations, and the computational power of traditional computing chips cannot meet the requirements of D** computing. AI chips with high computing power can meet the computing needs of AI industry and have been developed rapidly.
At present, because the computing power of terminal equipment is generally limited, model training and inference are mostly completed on the cloud server. In cloud model training, *VIDIA's GPU dominates, and multi-gpu parallel architecture is a common infrastructure scheme for cloud training. In cloud identification, based on the consideration of power consumption and computing speed, the gpu-based method alone is not the optimal solution. Taking advantage of the respective advantages of CPU, GPU, FPGA and ASIC, heterogeneous computing (CPU+GPU+FPGA/ASIC) is the current mainstream solution.
In applications such as computer vision and speech recognition, the cloud computing that the terminal collects data (especially image data) and then uploads it to the cloud for processing poses a growing challenge to network bandwidth and data center storage. In addition, applications such as unmanned driving require very high real-time performance and security. The delay and stability of the network brings about security risks that can not be tolerated by unmanned applications. Collecting data in the terminal and completing data processing to provide Edge computi*g of intelligent terminal inference has attracted more and more attention due to its ability to meet real-time and security requirements and save bandwidth and storage. The industry has decided that more and more of it will be done on terminal devices, that is, intelligence will sink down to terminal devices, and intelligent edge computing will rise.
Real-time is one of the main reasons for choosing to complete the inference at the terminal. However, due to the large number of parameters in the deep neural network, the process of inference needs to complete a lot of calculation, which puts forward high requirements on the computing power of the terminal hardware. In addition, battery-powered terminal equipment also has high requirements for power consumption, and most terminal products are price sensitive. That is to say, the terminal chip that performs D** deduction has strict limits on computing power, power consumption and price.
The embedded AI computing power chip developed by AGM integrates several technologies that have been used and verified by mass production in its existing FPGA+MCU products:
* ultra-low cost and low power consumption reconfigurable logic custom ASIC
* computing unit ASIC customization technology -- ASIC matrix calculation
* embedded memory read ASIC technology - low latency and low power consumption
* integrate the McU-led chip architecture and software ecosystem
The MCU computing power chip introduced by AGM is similar to ASIC in cost, but belongs to reconfigurable ASIC in architecture. It is suitable for AI high computational power requirements and constantly evolving algorithm upgrading, such as C**, R** and matrix operation. The chip meets the need for low power consumption, especially for low latency memory access.
This computing power chip of AI reasoning at the edge of Internet of things, featuring ultra-low price and high cost performance, can be applied in the terminal of the following application scenarios:
* AI analysis of edge camera images;
* voice AI analysis of smart speakers;
* image AI analysis of driving records;
* data AI analysis of industrial control;
* sensor data AI analysis
*AI high computational power requirements and constantly evolving algorithms C**,R**, and matrix operations
For example, in the security field, compared with the traditional video monitoring and AI+ video monitoring, the main change is to change the passive monitoring into active analysis and early warning, thus solving the problem of requiring manual processing of massive monitoring data. The terminal equipment such as security and unmanned aerial vehicle has a high demand for computing power and cost. With the development of image recognition and hardware technology, the conditions of intelligent security in terminal are becoming more and more mature.
In the automobile industry, real-time performance is the primary prerequisite to ensure safety in the case of high-speed driving. Cloud computing cannot guarantee real-time performance. The onboard terminal computing platform is the future of autonomous driving computing. In addition, with the development trend of electrification, low power consumption becomes more and more important to the automobile industry. The natural ASIC chip that can meet the real-time and low power consumption will be the future development trend of on-board computing platform.
In 2016, the global market size of AI chip was usd 2.388 billion, and some institutions predicted that the global market size of AI chip will reach usd 14.616 billion (the market size of terminal AI chip) by 2020, with huge space for development. In the future "foggy computing" terminal products, we believe that the low-price and low-power MCU computing power acceleration SoC chip will become the dominant product of AI terminal products in the future, and AGM should have a place.
AGM Micro has dabbled in AI terminal chips