In this GTC Taiwan, NVIDIA CEO is yellow future of Ren Xun period inside 10 years, will grow every year to the dimensions of operation demand 100 times, anticipate at the same time in rub Er is mensurable gradually under attenuation, the GPU operation quantity of before the whole world 50 large super computer will be in future grows inside 5 years 15 fold, the means that quickens operation with GPU at the same time will be become extend rub the main mode of Er law.

Huang Renxun emphasizes past NVIDIA creating place of CUDA operation mode to drive again in GTC Taiwan quicken benefit, explain future borrows the mode that quickens operation by GPU to will expand continuously at the same time, anticipate 2028 global operation demand will be equal 10 million groups of VoltaFrameworkGPU place promotes efficiency, if pass pile of much group CPU to make ability of operation of super computer order and degree with the tradition, will hold large-scale space and expenditure of electric power of high specified number, be like the word that replaces with GPU, can save more space and power loss, bring at the same time accelerate the effect higher.

Base of current and super computer develops main tool via becoming modern science, build chemistry of compose, quanta, quanta in the element respectively analysis of imitate of exploration of research of mechanical, weather forecast, atmosphere, the sources of energy, physics, data and development of artificial intelligence technology act important role, and be offerred two hundred and seventy-six million four hundred and forty-seven thousand two hundred and thirty-two times or efficiency of operation of one billion four hundred and ten million and sixty-five thousand four hundred and eight grade. And show with respect to OpenAI statistic, future the model of artificial intelligence operation inside 5 years will grow 300 thousand times, compare rub Er law anticipates growing rate is 30 thousand times rapidder, borrow quicken ability general to be able to make form of data, figure complex spend by GPU promote considerably, borrow this to solve associate with manpower inextricability operation demand.

Announced to roll out design of integrated Tensor Core last year, and the Volta of memory of integrated 32GB HBM2FrameworkGPU, lend operation of this corresponding 125 Tensor TFLOPS efficiency, part corresponding 7.5 FP64 TFLOPS or efficiency of budget of 15 FP32 TFLOPS, use GPU to quicken operation mode to be able to promote 10 times efficiency than associate with, let take up space and power loss are reduced considerably further at the same time.

And restrict to break through framework of strong put oneself in another’s position, NVIDIA announces to roll out NVSwitch further in the GTC 2018 this year, make 16 groups of Volta GPU can common be as high as 512GB HBM2 memory (32GB X 16) , add up to but corresponding efficiency of operation of core of 81920 groups of CUDA, 2000 Tensor Core TFLOPS, make the whole world the GPU of highest efficiency, and do not suffer GPU of limitation of traditional CPU framework access memory capacity influence. Borrow by the design of NVSwitch, NVIDIA more announce to roll out the whole world the biggest (and but amuse oneself game) DGX-2 GPU, correspondence borrows by be as high as 2PFLOPS operation efficiency, the machine case that and special and poriferous fiber is designed lets run power to be as high as 10000W maintains microtherm to run, the rolls out formally DGX-1 operation efficiency before comparing half an year promotes 10 times.

Must borrow than associate with form by CPU of 300 groups of double core, the server that must use up 180000W power specific power consumption to run, pass only group DGX-2 GPU can efficiency of corresponding and identical operation, but whole forThe priceNeed 1/8 and 1/18 power specific power consumption only, drill than associate with Alex Alex Krizhevsky passes two pieces of NVIDIA GTX 580 GPU at the same time, expenditure finishs training AlexNet between 6 climate, borrow need 18 minutes to be able to be finished only by DGX-2 GPU. At the same time DGX-2 GPU also breaks every second to analyse 1075 respectively video, make career of operation of the fastest odd chip, and every second can handle 15500 in every node video, and can be inside 14 minutes complete extension, inference defer time is in only 1.1 millisecond, every second more but inference figure 6250 video.

The operation ability that passes DGX-2 and NVSwitch are strung together receive a technology, NVIDIA also announces to roll out the server platform that builds form with DGX-2 to design HGX-2, and with wide amount to, Yun Da, Fuji health, flower course of study is amounted to, abb is achieved, abb glume, Hua Shuo, ability fine, the Taiwan such as Hua Qing, Tai’an, grand Chang cooperates in ground manufacturer, emphasize the whole world about 90% servers come from Taiwan at the same time, and NVIDIA also lasts in ground manufacturer with more Taiwan collaboration.

Borrow by GPU operation ability, cooperate the video processing technology that cooperates with the software firm such as Adobe, will can come true to be revised immediately video in needless thing, or it is to rebuild video in devoid content, can appear further even ” United States colour ” the effect. Pass the Kubernetes container group that offers with Google to administer systematic cooperation at the same time, will can let system of more artificial intelligence can differ because of answering operation demand trends adjusts operation efficiency, borrow this to let GPU wear fast operation efficiency to have stretcher configuration beneficial result, will cooperate with the manufacturer such as Alibaba, Baidu, EBay, HIKVISION, IBM, millet.

Be in with Taiwan collaboration part, NVIDIA expresses at present Fuji health will borrow detect by artificial intelligence technology makeProductionEfficiency, hospital of have as an attached institution of Chinese medicine university assists doctor analysis to forecast cancer tumor to change a condition through artificial intelligence technology, taiwan university distinguishs through artificial intelligence nose pharynx cancer endangers an organ, and lab of Taiwan artificial intelligence also passes artificial intelligence technology to assist the municipal government austral the stage to monitor bridgeStructurePrecautionary typhoon is damaged, the fair vehicle that peach garden municipal government plans to allow line of 30% fixed travel before 2020 can configure Level 3 to drive automatically function.

As be before during GTC 2018 with ” PLASTER ” regard a theme as speech ending, huang Renxun also emphasizes borrowing respectively by but process designing (Programmability) , low defer (Latency) , high spirit is spent definitely (Accuracy) , dimensions is changed (handling capacity of Size) , data (efficiency of Throughput) , specific power consumption (Energy Efficiency) , drive study to train efficiency then (Rate Of Learning) , let artificial intelligence can grow with rapidder rate.