Artificial intelligence technology is swift and violent the instantly of development, should come true more efficient, more the AI that human nature changes, of need is framework, algorithmic not just, return the cornerstone that helpful AI shapes, namely AI chip. Regard artificial intelligence as the carrier that is able to run, the development of AI chip is controlling the progress of artificial intelligence technology undoubtedly. Recently, china the strategy that it is AI is whole exposure, numerous AI chip appears formally, its driving function makes everybody exclaim when, also can’t help letting a person generate doubt, why China wish to want to transform oneself group for this scene, why does AI chip get of so much chase after hold in both hands?
Initiative phase, because do not need to consider the matter such as power comsumption, the chip of artificial intelligence can be gone after calculate force, high intercurrent, high handling capacity high, the solution that shows level is the different compose mode with “GPU+CPU” will finish commonly. In the meantime, this kind of plan basically faces the training link of each big AI enterprise and lab, be in as a result of AMD in the market at present of this respect be short of for a long time break, bring about deepness study GPU to quicken the market by NVIDIA alone big.
But develop application in high in the clouds gradually as artificial intelligence, raised more requirement to AI chip, while give attention to two or morethings is high-powered, still need power comsumption of give attention to two or morethings and reaction time. And regard half custom-built special integrated circuit as FPGA(spot but array of process designing door) can satisfy a requirement fitly, its characteristic is specific power consumption low, have the character that takes in and send out in large quantities small delay time, high at the same time, can satisfy the requirement that artificial intelligence runs in high in the clouds. Current, FPGA market Xilinx and Altera hold absolutely monopoly position twice, market share is adjacent 90% .
Future, artificial intelligence technology is destined to want farther transfer to a lower level to terminal unit, because physical reason suffers,be restricted, accordingly the power comsumption of chip, area, cost needs to be optimized further. Main solution is independent ASIC and SoC+IP independent ASIC of two kinds of; can be passed custom-built the design that change will achieve performance the chip with more advantageous, confidential more powerful sex, but defect is development cycle longer, devoted cost is huge, average company stands hard. SoC+IP has very large dominant position on the development corrupt practice of ASIC, but defect is functional develop malleability weaker, before this China vacate series AI chip to be developed according to this kind of plan for issuance .
Although ASIC has the difficulty on a variety of development, but as special integrated circuit, the advantage of its function and power comsumption is clearly however. In computation of network of artificial intelligence nerve, have certain distinction with traditional computation, bring about when undertaking nerve network is calculated, CPU, DSP, GPU has calculate force, function, can the inadequacy of the respect such as effect, arose to design the demand of processor or accelerator for nerve network computation only so. Because ASIC is belonged to custom-built kind chip, involve redesign circuit, the difficult sex that produces accordingly is quite big, at present bibcock of domain of ASIC of the artificial intelligence in the market should belong to Google TPU, double what the function of the itself of processor of the 3rd acting TPU that its released May this year is the 2nd generation.
It is with the mobile phone exemple, in the smartphone there are GPU and CPU commonly in this kind of equipment, can offer corresponding operation ability to support for artificial intelligence, why does that develop AI chip even technically to have this job? Here cites a simple case, go all sorts of course can choosing when common restaurant has a meal is, may compare delicate, but if want to sample authentic dish, should go special restaurant just eats likely.
AI chip also is such, although the mobile phone also can have GPU and CPU, but power comsumption low, weight light, ply is thin just be the way that its seek. Appropriative AI chip can realize highest efficiency and ability, and take up very little space reachs lower power comsumption only. Won’t take up CPU reachs the resource with too much GPU at the same time, can assure to other operation still can have while the mobile phone is running artificial intelligence.
Of course the importance of AI chip is more than hereat. It is with Google TPU1 exemple, its have the pulsatile array of 256X256 dimension, it is 700MHz about, have 64K multiplication is unit, every unit Chancike carries out a multiplication and addition, namely 128K operation. So the frequency of summit summit computation of TPU1 every second is make an appointment with 90Tops, of course here is theoretical numerical value only, because data is transmitted, the reason such as memory, extraction, often rate of short of peak value.
But, relatively will undertake training at common GPU and CPU plan when, especially nerve network size is even more at present enormous, parameter is more multifarious, normally the training time of large NN model is as long as a few weeks even a few months, and during appear the plan that all sorts of accident circumstances also can affect training, wait like power cut. And it is OK to used TPU1 finish inside the time of a meal, its efficiency promotes greatly.
And AI chip although by the coronal the name with AI, but its essence or a chip. After the engage in a battle that experiences initial stage of Sino-US trade war, domestic company has waked up to reality in succession the core technology that if do not have oneself,comes over, the sculpture that although accomplish,also piles a beach again by force to go up nevertheless just, tide can complete collapse.
In face CPU and GPU to be issued by the pattern of foreign business monopoly, AI chip became the breakthrough with domestic new firm to nod. Will look at present, sino-US two countries is in artificial intelligence technology basically to go up with one scratch line, because this is more advantageous to the development of AI chip. But the result that the design of chip is not in one day, the chip engineering capability of together with our country is very fragile also, need throws financial capacity of much labor power to just may reach effective product. The company that home throws research to AI chip at present has The Cambrian, Hua Weihai to think of, elder brother of the crop in resurgence microelectronics, A, still a lot of small business also follow suit in succession in entering AI chip field.
Because AI chip makes different tradition,this is chip is general, do not need to cost energy to make IP kernel, use other IP directly plus framework level optimize, undertake conformity to IP in the light of demand for service, the doorsill that makes accordingly is reduced greatly, as above relates mentioned SoC+IP plan. China roll out for what vacates series for infuse of AI homebred chip stitch cardiac, greatly hearten the company is created related homebred AI chip.
But notable is, although China to rolled out AI chip, but the application that its carry in the mobile phone still has a lot of limitation, apply on picture identifying only mostly. Not only China be so, apple and couplet hair division also are such, how to extend the more application of prospective AI, became the key that will stride next.