news

interface news reporter | cui peng

interface news editor | song jianan

9moon25daymorning，baiducloud intelligenceat the conferenceroll outbai geAIheterogeneous computingplatform4.0andqianfan large modelplatform3.0waitAIinfrastructure products，and announcebemultipleAIrelated businessofup to datedata。in,literary mindlarge modeldaily averagecallquantityalreadyexceed7100 million，distancebaidulast timeannouncementof6100 milliondatahavefurtherpromote。

existjust endedaliyunqi conferencesuperior，alibabaCEOwu yongmingeverexpress，alibaba cloudofsingle networkclusteralreadyexpand to10ten thousandcardlevel,andbaidu alsounwilling tobehind。shen dou, executive vice president of baidu group and president of baidu intelligent cloud business groupspecialemphasize，bai ge4.0willnot onlyyeswankacluster，ratherhavemature10ten thousandcardclusterdeployandmanagement capabilities。

baiduthis yearalwaysemphasizelarge modeloflandingapplicationbaidu ceo robin li in a recent internal speechindicated in，haveapplication scenario、cancontinuous iterationupgradeoflarge model,andothermodelthe gap between products willgetting bigger and bigger.

based on this,baidufound itchangan automobileandsamsungwaitlargecustomers stand for them,byexhibitthe application results of baidu's big model in various industries.

shen douexpress，in the past year,baidufeelclientofmodelneedsurge，the required cluster size is getting larger and larger.enterpriserightmodelinference costdeclineofexpectedalsogetting higher。

training large modelsthe premise iscreatecluster, this is not just about buying a gpu and assembling it.generallyneeda few monthsconductequipmentconfigurationanddebug。

previously there wascloud vendorsmentioned, forming a clustercancompressionarrive1sky,andshen dousay，bai ge4.0be able to dofastest1hours to completeformation,mainusewillpopular training in the industrytoolandframeconductbuilt-in way.

onceenterlarge-scale trainingstage，most importantthat isstability。the field of large models has always followed the famous scaling law, which believes that model performance will improve as the parameters, computing power, and data set size increase.。

according to shen dou, gPUthe cluster needs to consumehugeofconstructionand operating costs, usually builtonewankacluster，oneyesGPUofpurchasecostat oncegundambillionsyuan.ifyes10ten thousandcardcluster，serverone dayconsumptionofpoweraboutyes300ten thousandkwh，equivalent tobeijingdongcheng districtone dayofresidential electricity consumption。

existthislarge scaleofon the cluster，hardwareinevitablemeetingappearfault，the bigger the scale，faultyprobabilitythe higher。existthesefailure，the vast majorityyesdepend onGPUcaused by，becauseGPUyesvery sensitivehardware，righttemperature、humiditywaitenvironmentfluctuationresponsive。

shen dou mentioned，when meta trained the llama3 model, a cluster of 16,000 gpu cards was used, and a failure occurred on average once every 3 hours.

large modeltrainyeshugesingle task，onenodeerror，entireclusterjust needstop,androllbackarrivepreviousmemory point。consideringGPUclusterofcostveryexpensive，everymanystopone minutewillwaste of money，“effective trainingduration"it becomesveryimportant indicators。

againstlarge modeltrainin processfrequent failuresofquestion，bai ge4.0rightfault detectionmeansandautomatic fault tolerance mechanismconductbeupgrade，at presentwankaclustersuperiorofeffective training timeachieve99.5%，shen dousaythis is higher thanpeersopponentofdataperformance.also,bai ge4.0willmainstreamlong textinference efficiencypromotebe1timesabove，at the same timereducedinference cost。

in baidu's latest earnings call,robin lieverit was revealed that in the second quarter, the revenue contribution of baidu smart cloud ai further increased to 9%, compared with 6.9% in the previous quarter.

large modeltoolofperformancepromoteof courseimportant，but forbaiduto say，model landingofresultsmore realistic。

in addition to upgradingbai gein addition to the platform, baidu also introducedlatestqianfan3.0platform. according to the data released by shen dou,on the qianfan big model platform, wenxin big model is called more than 700 million times per day on average, and has helped users fine-tune 30,000 big models and develop more than 700,000 enterprise-level applications.qianfan 3.0 canit can call nearly 100 large models from home and abroad, including the wenxin series of large models, and also supports the calling of various traditional small models such as voice and vision.

at present,large modellandingofthree majormain requirementsthey areapplication development、modelreasoningandmodeldevelopment。

existapplication development layer，enterpriseRAG（willenterpriseandindustry datamakehacksknowledge basegivelarge model）andAgentagentyestwo majorcommonoflarge modellandingscenario。

after receiving the task, the agent will think independently, break down the task, plan the solution, and call the tools to complete the whole process autonomously.，cancomplete the pastneed3 to4indivualAPPtalentfinishedtask。

shen douexpress，baiduinternalforqianfanplatformsupplybebaidu searchsearchandbaidu mapwaitexceed80indivualofficial components，used forpromoteagentexiston specific tasksability。

and inin li yanhong’s view, intelligent agents are large models.developmentalnextimportant direction."many people are optimistic about the development direction of intelligent agents, but to date, there is still no consensus on intelligent agents. there are not many companies like baidu that regard intelligent agents as the most important strategy and development direction of large models."

baidu released three products at this year's create conference, namely agentbuilder, appbuilder and modelbuilder. among them, agentbuilder and appbuilder are both related to intelligent agents, one has a lower threshold, and the other emphasizes functionality.

according to the latest data revealed by baidu, the distribution volume of intelligent bodies in the baidu ecosystem has increased significantly, with the average daily distribution times exceeding 8 million in july, which is twice the data in may.

baidudigital humanplatformandsmart customer service products are also availablelatest developments。among them, xiling digital human platform 4.0 supports the rapid generation of 3d digital human images and videos with different makeup and industry characteristics based on text, and reduces the price of 3d hyper-realistic digital humans from 10,000 yuan to 199 yuan.

the intelligent customer service product "keyue" has been optimized in terms of user intent understanding and multimodal information exchange, improving its ability to handle complex problems.

according to baidu, the industry“problem self-resolution rate”ofthe average level is 80%, and after the upgrade, "keyue" has raised this indicator to 92%. the product has helped corporate customers serve more than 150 million people and interacted more than 500 million times.