Artificial intelligence (AI) is the technical development trend with the most fundamental in last few years industry of science and technology, not merely chip industry person in succession layout, a lot of application development business also are guiding AI hard oneselfProduct. Nevertheless, to application development group character, AI still enters the technology with very high threshold at present, and relevant development environment still has very big improvement space.

If where of oneselfProductArtificial intelligence element is added in, it is the problem that a lot of application development dealer are thinking. Conserve besides intelligent sound box, family expenses outside the consumptive sex product such as cinematograph, monitoring of condition of stage of vision of machine arm, machine, machine, even semiconductorEquipmentWait for the application development business of industrial domain, also guide artificial intelligence in the attempt oneself product, in order to create more and additional value.

However, the development of artificial intelligence application, to a lot of developer, still put in extremely high threshold. The application development of artificial intelligence can be divided into ” the cloud ” with ” upright ” two much. What the cloud points to is center of data of high in the clouds or super computer, have powerful operation capacity, responsible executive model trains the job; End is to show good with training model undertakes deductive all sorts of brim nodeEquipment. Correspondence uses developer and character, the model trains the start that is artificial intelligence application development, also be the germ of a lot of problems.

TraditionalHardwareTraining efficiency is low HPC still too high to reach

The training of AI model, it is the task of special arduous to operation equipment. Arrive to get more the identifying with accurate essence result, what AI performs algorithm is complex degree, data precision grows continuously, the operation efficacy requirement that model training needs also follows cruel add, the servo that is a foundation with CPU to the tradition implement, super computer brings very big challenge.

NVIDIA technologyBe on saleManager Su Jiaxing (graph 1) points out, as rub Er is mensurable (delay of hasten of Moore’s Law) feed rate, the efficiency of CPU grows slower and slower also. Sheet relies on the efficiency that CPU place provides, already did not follow to go up the demand of deepness study application development. And this also is GPU operation be in the last few years is wide the reason that catchs each attention- – the super computer that is a foundation with GPU, can provide taller operation efficiency, to be engaged in the need such as training of model of advanced science research, artificial intelligence the group of research and development of ability of extremely expensive operation is used.

Graph technology of 1 　 NVIDIABe on saleManager Su Jiaxing points out, the super computer that is a foundation with GPU has efficiency of more expensive operation, can satisfy AI model to train place to need.

It is with NVIDIA exemple, this company besides chip of GPU of research and development, have user of extremely high demand to operation efficiency to offer complete solution to give, still roll out super computer of the DGX that is core with GPU and DGX-2, use high guest to make the system that changeFrameworkWith in-house interconnection technology, in order to break the efficiency bottleneck of traditional and super computer, and whole is hadCostFar superer than the tradition computer comes cheaply.

With by servo of CPU of 600 double core implement the crowd together that forms (Cluster) is exemple, so huge system, need at least tens of machine ark just is put, power consumption is as high as 360kW; The servo that uses 4 GPU of Tesla V100 to configure implement, need 30 only, can provide coequal operation efficiency, HardwareCostIt is about 1/5 former, power comsumption and the space that take up have 1/7 former only. If main demand of the client is to undertake deepness studies research and development, can consider to replace whole crowd together with a DGX-2, hardware cost needs only 1/8, power comsumption is reduced 1/18, take up the space has 1/60 of traditional crowd together only more.

Nevertheless, even if DGX has let super computer with DGX-2 is large company no longer even the equipment that ability of center of national level calculation cans afford, the enterprise to a lot of Taiwan, of DGX seriesThe priceToo costly still, the burden does not rise. The person that the stage ties line of business discloses, the official list price with current DGX is 69 thousand dollar, be equivalent to new station money 2 million yuan of fluctuation; But the quote of DGX-2 is new station money at least 1, 0 yuan take off. To a lot of manufacturers, brushstroke persuades decision-making high level very hard to buy sheet thisPurchaseDemand. Accordingly, the group of research and development of this company still uses the servo that inserted much Zhang Xianshi to get stuck implement will train AI model. Although training rate is very slow, but at least cost of research and development won’t exceed a budget.

TallCharacterData is obtained not easy

Besides tall efficiency hardwareThe priceOn the high side, many station plants can outside indigenous method steel-making, the data set that training uses is obtained not easy, also be Taiwan course of study person the one catastrophe problem that development encounters on AI application. When undertaking the model trains, of data setCharacterCan affect the result that the model trains directly, but the collection of data set follows build place, itself is a drudgery, if want to build place high quality, cover whole data set, more need to throw material resources of much labor power. Detect with the product used machine vision system is exemple, if the data set on developer hand is insufficient complete, when training the flaw that the model that come out does not have in encountering data set, taste probably for fine with respect to meeting miscarriage of justice. However, product flaw basically is normal distributings, most flaw is the flaw of a few kinds of a few specific types, certain and infrequent flaw is likely thousands of, even tens of thousands of data just appears.

This kind of infrequent flaw, it is data set builds the state with the most intractable buy. Developer must spend a long time to just the opportunity collects this kind of data, avoid a model to produce a miscarriage of justice when inference then. In fact, all sorts of artificial intelligence application have similar problem, when the unusual situation of few number happens, the judgement of artificial intelligence appears very easily error.

Nevertheless, this kind of problem also is not cannot can solve. Be able to bear or endure canSoftwareDesign manager Shen Mingfeng to point out, in the process that trains in the model, moderate land adds artificial noise in data set, conduce to the accuracy rate that raises model judgement. This company uses this kind of technology to train a model namely.

Not come singly but in pairs, the endowment plan that training platform in model of AI of development high in the clouds is met transition place, also develop the discovery in the process in platform, right video undertake gyral, screwy, can let a model be when training, even if uses lesser data set only, also can get similar the training result that uses complete data. In addition, at present industry is supervising the technology such as type study, move type study in development blame, depend on in order to reduce what logarithm occupies part.

The guest is made change the model often won’ts do

Be worth what carry is, the engineer of experience of a lot of development having AI is mentioned, although had had the nerve network model of very much open and primitive code at present, but to satisfy requirement of certain and special application, these off-the-peg models or the settle or live in a strange place that need to pass certain level are made change revise.

However, make in classics passing traveller change after revising, the circumstance that these models cannot carry out often happens. Shen Mingfeng says frankly, this is the problem that a lot of clients are developing AI application to be encountered constantly, he himself has treated a few case. Sometimes the problem is to go out in tool itself, it is the modification range that the client makes to model place occasionally too great place is brought about.

Undertake in the light of the model the guest is made change revise, can assist application development business to realize product diversity to change, or oneself alone Know-how puts the door into the product in, because this guest is made,turn a model to a lot of application development business character is very pivotal, and it is absolutely that relevant model follows data set cannot the trade of outflow is secret, even if wants to entrust a partner to assist debug, it is the job that produces unlikelily.

Accordingly, a mature zoology department and industrial standard, to AI application development comes for the key, and this also is An Mou (the end that Arm) urges Project Trillium. An Mou is seniorThe marketIan Smythe of chief inspector of be on sale expresses, this company published Project Trillium in Feburary 2018, include what can move in Arm processor among themSoftwareCase type library, learn to run a machine only (ML) exercise adds the machine study processor that carry, and let other AI processor be able to join the function of Arm processor system.

In Project Trillium, the nerve network that An Mou defined him (2) of NN) SDK(graph, the nerve network of all sorts of industry main trends carried on on itsFrame, for example TensorFlow, Caffe, its lower level is operation case type library, it is hardware ground floor next. Be worth what carry is, this SDK can assist Cortex-A not merely processor of this kind of tall efficiency, be aimed at the small controller that is core with Cortex-M series (MCU) , also have corresponding CMSIS-NN.

Graph 2 　 Arm NN SDKFrameworkSketch map

Industry public figure thinks, of Project Trillium roll out, indication Arm uses him to design the dominant place of the domain in IC of purpose, appear personally body of the hardware that all adds up to contention of a hundred schools of thought, pliable but strong is rock-bottom. This follows the consistence between hardware to increasing a model, also can help somewhat. If Arm NN can be become wide the mainstream level that captures industry to accept, when prospective industry or academia are developing new model, just imagine this model future along with all the others likely Port is in the likelihood on which kinds of hardware, try to reduce consistence problem in the beginning of design thereby.

Application development brings AI example changes style concussion

AI brings example move for industry of science and technology, but opposite also overturned the environment of be accustomed to sth follows everybody work habit. Besides IC design firm, who can think of one day oneself company to develop a product, get the operation efficiency that uses super computer social estate. As to the consistence problem of the model, for the staff of research and development that makes equipment of PC, action to the habit, most propbably also is the thing that thinks very hard before, because X86 follows Android,basically had died applied casing. However, in AI be in power today, had done not have what is of course it seems that.