NVIDIA author held presiding apparitor Huang Renxun concurrently to release HGX-2 on the Taipei GTC that held recently, this is " compose builds module " cloud server platform, let server manufacturer can establish more powerful system around NVIDIA GPU, realize high-powered computation and artificial intelligence.
The performance of this platform grew 500 inside 5 years times, got inclusive all computer manufacturer and ISV inside the support of ecosystem.
Huang Renxun releases HGX-2
Huang Renxun expresses, computational system brings the industry to the dollar is counted inside global limits the change, and GPU is in core position in computational ecosystem. He thinks to want to come true to wait for a series of breakthroughs of the respect in material science, the sources of energy and medicine, the computational capability with higher hang on.
"Of computational demand big unprecedented, we need to expand rub Er is mensurable, " Huang Renxun expresses.
Function violent wind risesHuang Renxun is expatiatory the technology of the deepness study drive that is based on GPU " The Cambrian explodes greatly " . Inside the time that does not arrive 10 years, GPU computation ability grew 20 times, grow 1.7 every year namely times, yuan Chao rub Er is mensurable.
But the development as a result of AI, we still are in to the demand of function " growth (and did not put delay) " . "Before this, software is written by the mankind, software engineer can write finite software only, but the machine won't feel exhaustion. But the machine won't feel exhaustion..
"Want to have data only, want to know how to establish a framework only, we can make powerful software, " Huang Renxun says so. "The whole world develops the company of software every to need an AI super computer. "The whole world develops the company of software every to need an AI super computer..
Roll out brand-new serverNVIDIA HGX-2 is shirt-sleeve the revolutionary function such as structure of NVIDIA NVSwitch interconnection, do not have 16 NVIDIA Tesla V100 Tensor Core GPU seam join to be onefold " giant GPU " . The partner of NVIDIA will roll out first system that is based on HGX-2 later on this year.
"Had the collaboration with you, anybody can use means of confluence of this kind of future computation to realize what high-powered computation and AI apply to found, " Huang Renxun expressed to thank to the support of the partner in whole computer industry. "We have all sorts of servers. "We have all sorts of servers..
Trade partner adopts the server platform of NVIDIA extensively
Big everlastingThe core of HGX-2 is NVIDIA Tesla V100 GPU, it provided capacity of memory of 32GB tall bandwidth, can offer every second 125 trillion times floating-point calculative deepness studies performance.
Pass NVSwitch to will amount to 16 Tesla V100 GPU to join together, in bringing up Huang Renxun mouth " the GPU with the greatest whole world " .
"Every GPU can have communication with the bandwidth of 300GB/s, PCI Express of 10 times and other GPU at the same time, " Huang Renxun says. "That is to say all GPU can speak at the same time between each other. "That is to say all GPU can speak at the same time between each other..
Huang Renxun still introduced new fund NVIDIA DGX-2 in detail, this is the system that compose of platform of server of first use HGX-2 builds. It hefts 350 pounds, can offer every second 100 million times floating-point calculates 2 ten million ability and 512GB HBM2 memory.
"This is the fastest onefold computer that throughout history founds: Model of an an operating system, process designing, you can be its process designing a computer, " Huang Renxun say. "This resembles is a personal computer, rate is quite rapid nevertheless. Rate is quite rapid nevertheless..
Industry new lawHuang Renxun expresses, DGX-2 came true with 6 months only decuple at going up generation is calculated the leap of function. In AlexNet pattern recognition in fiducial test, relatively also realized the leap of 500 times at the performance that achieves to NVIDIA GPU place is based on 5 years ago.
"Industry had greeted new law, " Huang Renxun say. "This new numeration tells us: ' if you can be accomplished, and if you are willing to undertake in whole stack,optimize, the rate that achieves function promotion is met very fast. ' "
Its are as a result: The technology issued Dou Chong to show records of a series of deepness study speed on stack, reach giant data center from odd chip function system.
NVIDIA GPU Cloud (NGC) expanded this is planted of function can visit a gender. It lets researcher get machine of table of be confined to and server system no longer, can lend force the cloud system that provides at waiting for a provider by Yamaxun, Gu Ge, Alibaba and inscriptions on bones or tortoise shells of the Shang Dynasty however.
"Each of this one software actor or actress through tone and passed a test, " Huang Renxun returns complement to say, 20 thousand companies had downloaded NGC nowadays.
NVIDIA provided record-breaking performance up and down for technical stack
The result of real world makes a person exclaim. Huang Renxun is attendant go up with a NVIDIA AI researcher together, demonstrated to move in real time the detail in dividing a photograph, for example tree or street lamp.
Remember PLASTER principle wellHuang Renxun say, besides demonstrate besides, want large-scale deploy such deepness study application, need does decide 7 big challenges: But process designing sex (Programmability) , defer (Latency) , accuracy (size of Accuracy) , model (Size) , handling capacity (Throughput) , can effect (Energy Efficiency) and study frequency (Rate Of Learning) . The English initial that uses these 7 words is PLASTER for short.
The explanation says Huang Renxun, inference is not easy thing.
Patulous ability will become service of artificial intelligence of will new generation to push to the key that runs actually, this is a kind of procedure that calls inference, apply to speech synthesis and identify, image and video processing, and recommend a service to wait a moment.
Wang Zhe " KONG "To show this kind of patulous ability, huang Renxun demonstrated flowers of an exclamatory making a person to identify a system, the speed that is based on a CPU to be able to be achieved pursues 4 pieces for every second picture, and use a GPU to be able to expand to every second 2500 pieces pursue picture, 4 times what still can achieve this one rate more highest, this should be attributed to the ability that is located in center of desktop, data and cloud service provider to add more GPU to support in order to offer quickly for more through Kubernetes.
"The goes up for NVIDIA GPU Kubernetes (Kubernetes On NVIDIA GPUs) that we weigh, abbreviation KONG, " Huang Renxun say.
Kubernetes of have the aid of, what GPU computation can achieve your place to need is the most large-scale
"Kubernetes it may be said exceeds massive operating system, " Huang Renxun say. "If you visit whole warehouse, arrive from GPU can put Docker and recipient all medium, API that move on Kubernetes and library, stack of this one software is so complex, the job that hundreds engineer comes to several years is around move it. The job that hundreds engineer comes to several years is around move it..
New GPU, new marketNVIDIA server offers the trade that counts a dollar for value exceed strong computational capacity.
In value the recreational industry of 2 trillion dollar, NVIDIA is new the RTX technology that roll out can quicken the light to track, provide the standard of film level quality, plus technology of real time graph and artificial intelligence, the traditional apply colours to a drawing that will be based on CPU now quickly farm. In value in the Medical Protection industry of 7 trillion dollar, our Project Clara hopeful comes through GPU computation redefine medical treatment is become picture, be based on current medical treatment instrument to realize taller fidelity, or the image that uses fewer ray a future life to become same quality (the ray of scanner the quantity decreases 1/6, can use at children surely) . Safe city also represented another value the market opportunity of 2 trillion dollar. The DRIVE platform that NVIDIA is devoting oneself to to carry end through its solves traffic to carry the problem of the domain. Traffic carries a delegate to wear another value the market of 10 trillion dollar. "Did not come to the thing that everything can move will be automation, " Huang Renxun introduced billions of miles to NVIDIA is being collected and handle model of data, training, imitate in detail drive, and the ability that drives truly. "Contractible " and twinkling moves in miniature spaceIn speech end, technology of virtual reality of Huang Renxun have the aid of works in the same place ours " contractible " and twinkling moved in car of a miniature, next this colleague is opening this car travel in city of a miniature.
That is to say, had such technology, the mankind can become the mothball force of AI machine in fictitious space, no matter these machines are located in where, size how.
"Future, your aux will be able to is enough with the robot ' be in harmony is an organic whole ' , " Huang Renxun issues official of many 2000 developer, researcher, government and media say to the stage, "Enough implementation of your aux will be able to is long-range and extemporaneous (Telepresence) , reach any places that want to go. Reach any places that want to go..
