Classification and application of AI chips

1. Classified by technical route

at present , AI chips are divided according to the technical route , It is mainly divided into GPU、FPGA as well as ASIC Three broad categories , The characteristics of the three types of chips are shown in the following table .

AI chip	Technical characteristics	advantage	Inferiority
GPU	GPU Called graphics processor , It is the core unit of the graphics card , Is a single instruction 、 Multiple data processors .GPU It adopts a large number of computing units and ultra long pipeline , It mainly deals with the operation acceleration in the field of image processing . The representative products are Nvidia-Tesla v100.	advantage ： comparison CPU, Due to its multi-threaded structure , Have strong parallel computing ability , And FPGA and ASIC comparison , More versatility .	Price to power ratio FPGA and ASIC higher , The ability of parallel operation cannot be fully exerted at the inference end .
FPGA	FPGA Called field programmable gate array , Users can repeat programming according to their own needs , Applicable to multiple instructions 、 Analysis of single data flow , And GPU contrary .FPGA It is often used in the prediction stage , Implement software algorithm with hardware , therefore , It is difficult to implement complex algorithms .	It can flexibly compile the hardware layer of the chip , And the power consumption is far less than CPU and GPU.	Hardware programming language is difficult to master , The computing power of a single unit is relatively weak , High hardware cost , Electronic tube redundancy , Power consumption can be further compressed .
ASIC	ASIC It is an integrated circuit designed for a specific purpose , It is a kind of special-purpose AI chip . Except that it cannot be extended , At power consumption 、 reliability 、 There are advantages in volume , Especially when high performance is required 、 Low power mobile terminals .	Optimize the design of architecture layer for special tasks , Can be realized PPA Optimal design 、 The cost is the lowest after mass production .	The initial design investment is large , Programmable architecture design is difficult , Targeted design will limit the versatility of the chip .

2. Classification by application field

Divide from application fields ,A Chips can be mainly divided into cloud data centers AI Chips and smart terminals AI chip ; And from the function to divide , It can also be divided into AI Training （Training） Chips and AI Reasoning （Inference） chip . The cloud market is basically NVIDIA and Google Two giants rule ,2020 Containing light developed by aridamo Academy in 800AI Chips have also entered the competitive ranks of cloud reasoning ; Terminal players are relatively more , Include Intel、 qualcomm 、ARM、Imagination、 Huawei 、 Cambrian, etc . 

AI There are many application scenarios for chips , It mainly includes ： Data Center (IDC)、 Mobile terminals 、 Intelligent security 、 Autopilot 、 Smart home, etc .

(1) Data Center

For cloud training and reasoning , At present, most of the training work is completed in the cloud . Video content audit of mobile Internet 、 Personalized recommendation is a typical cloud reasoning application .Nvidia GPU In terms of training, one company is dominant , Also maintain a leading position in reasoning . meanwhile ,FPGA and ASIC Because of the low power consumption 、 Low cost advantage , Is also continuing to snatch GPU Market share . At present, the representative chips in the cloud mainly include Nvidia-Tesla V100、Nvidia-Tesla T4、 Huawei shengteng 910、 The Cambrian MLU270、 Ali Dharma hall contains light 800 etc. .

(2) Mobile terminals

Mobile terminals mainly include mobile phones , And the increasingly popular unmanned aerial vehicles . Mobile terminals AI The chip is mainly used for mobile terminal reasoning , Solve the user experience problems caused by network delay in cloud reasoning , Typical applications such as video special effects 、 Voice assistant, etc . Mobile AI Reasoning is done on the system chip （system on chip,SoC） Add a coprocessor or special acceleration unit to realize . Subject to the power of the mobile terminal , There are strict restrictions on the power consumption of the chip . The representative chip has Apple A12 Neural Engine And Huawei Kirin 990.

(3) Intelligent security

The main task of intelligent security is video structuring . By adding AI chip , It can realize real-time response 、 Reduce bandwidth pressure , Besides , The reasoning function can also be integrated into edge server products , To realize the background of non intelligent camera data AI Reasoning .AI The chip needs to have the ability of video processing and decoding , The main consideration is the number of video channels that can be processed and the cost of single channel video structuring . Huawei is the representative chip Hi3559-AV100、 Haisi shengteng 310 And bitland BM1684 etc. .

(4) Autopilot

AI Chip as the brain of the car , Need a lot of on the car sensor The generated data is processed in real time , Computing power on the chip 、 Power waste 、 There are very high requirements for reliability , At the same time, the chip needs to meet the vehicle regulations and Standards , Therefore, the design is more difficult . At present, the chips for automatic driving mainly include Nvidia Orin、Xavier and Tesla Of FSD etc. .

(5) Smart home

stay AIoT Time , Every device in smart home needs to have certain perception 、 Inference and decision-making functions . In order to get a better user experience of intelligent voice interaction , voice AI Chips have entered the end-to-end market . voice AI Chip design is relatively low , Short development cycle . On behalf of the chip, think twice TH1520 He Yun Zhisheng swift UniOne etc. .