Giants turn to the internet of things, the next stop in the internet of Things is AI

650) this.width=650; "Src=" Http://s3.51cto.com/wyfs02/M02/85/A7/wKioL1erHiLwUKUBAACuttypEWE781.jpg-wh_500x0-wm_3 -wmp_4-s_1314795211.jpg "title=" deep-learning-header1.jpg "alt=" Wkiol1erhilwukubaacuttypewe781.jpg-wh_50 "/>

In July 2016, shocked by the world's shocking 23.4 billion pound takeover of the arm company, SoftBank chief executive Masayoshi Son said the acquisition marked a "paradigm shift" in SoftBank's investment in the Internet of things. It is true that arm, a chip design firm that has monopolized the smartphone industry, has been aggressively expanding into the internet of things for the last two years as the internet of things is poised to become a strategic opportunity for the size of the smartphone industry over the next few years.

However, ARM only provides the design of IoT chips, and even if Gartner predicts that there will be 26 billion IoT devices in 2020, this is just the physical foundation of the Internet of things. How to analyze, judge and trade the data of the 26 billion-device 7x24 flowing stream, it is not enough to meet this demand, but must rely on the automatic algorithm, this is Gartner actively advocated by the algorithmic economy.

AI will become the mainstream business competition strategy

Why is AI going to become a mainstream business competition strategy? This is because in the algorithmic economy era, artificial intelligence is the ultimate algorithm, the pursuit of the ultimate algorithm will become the mainstream business competition strategy.

In the 2015 book "The Ultimate Algorithm" by Professor Pedro Domingos of the University of Washington, "The ultimate algorithm" is the automatic discovery and creation of "main algorithms" for all other algorithms through machine learning. This so-called "main algorithm" is a single, global universal Algorithm, the "main algorithm" for the biological world is the human brain, for the internet of things-based machine is artificial intelligence.

Over the past 60 years, the world's top scientists have been studying how to simulate human intelligence in a mathematical way. Early proof of the "logic theorists" of the "Mathematical Principles" program to make the machine with logical reasoning ability, the medium-term expert system allows the machine to acquire human knowledge, and then to the machine to learn knowledge of the machine learning algorithm, scientists have not stopped the exploration of artificial intelligence.

Entering the 2016, also the 60 anniversary of the birth of AI, the machine learning algorithm based on deep learning becomes the mainstream of artificial intelligence, and the core of deep learning is the multi-layer deep neuron network DNN, which is the most mature machine intelligence that the mainstream scientists can reach at present.

July 21, 2016, Technology solutions provider SoftServe released the big Data Snapshot study, which shows that 62% of medium and large companies want to use machine learning for business analysis in the next two years. This means that the commercial multilayer deep neuron network DNN, will become the major enterprises chasing the mainstream business competition strategy.

Hardware Advancements drive AI business

We are entering the commercial era of artificial intelligence. After 60 years of development, artificial intelligence based on DNN has entered the enterprise from universities and laboratories, and spread from enterprise to tens of thousands of households. Google has just released this year's smart hardware Google Home, the production of Ali Internet cars, Microsoft AI Assistant Cortana, etc., are based on the dnn of large-scale commercial applications.

In the broader context of traditional enterprise applications, a large-scale artificial intelligence commercialization is beginning, which is based on hardware advancements, one notable improvement being the rise of the GPU in the artificial intelligence business. Wired magazine published an article entitled "Competing with Google, Facebook's open source AI hardware," in the end of 2015, referring to the rise of GPUs in the era of AI business. Now, Facebook, Google, Microsoft, Baidu and other large internet companies are turning to use the GPU to complete artificial intelligence business applications.

In the past, algorithmic models relied on CPU computing, but the uniqueness of AI algorithms lies in distributed parallel computing, which is not a CPU based on serial computing. Actually the GPU for graphics image processing is massively parallel from the start, which is why Professor Wunda of Stanford University will consider using the GPU to optimize the AI algorithm. Research shows that 12 NVIDIA GPUs can provide deep learning performance equivalent to 2,000 CPUs.

Although in the long term, the development of real AI chips is still a global academic and business needs to complete the task, but it still takes a long time to explore and experiment. From the current commercial demand for artificial intelligence, the use of the GPU for artificial intelligence is obviously a great advantage. That's why, in April 2016, Nvidia launched the world's first deep learning supercomputer DGX-1 based on GPUs.

First GPU deep learning supercomputer

The first GPU deep learning supercomputer nvidia DGX-1 is based on the Nvidia Tesla P100 GPU, which uses the latest NVIDIA Pascal GPU architecture in 2016. Pascal, the fifth-generation GPU architecture, was announced two years ago at the GPU Technology Conference (GTC), and the product will be available in 2016, the Tesla P100 GPU.

As a new generation of GPU architectures, Pascal has a significant performance boost compared to the previous generation of Maxwell. According to Nividia data, the Pascal GPU has a 1 order of magnitude improvement in the performance of training deep neural networks. At the 2015 GTC Conference, it took 25 hours to train the alexnet deep Neural network with 4 maxwelll GPUs, and 8 Pascal GPUs at the 2016 GTC Conference for 2 hours. Compared to the Intel dual Xeon E5 Server Training Alexnet Network takes 150 hours, while DGX-1 only takes 2 hours.

Alexnet Neural Network is the 2012 International imagenet Computer Graphics recognition Competition champion deep Learning algorithm, the famous open-source deep learning algorithm Caffe is based on Alexnet. And all of the best results in the ImageNet competition by 2015 were deep neural networks based on deep learning and GPU acceleration, and it's no wonder Wired magazine marveled at the rise of the GPU in the AI era.

The advantages of the Pascal GPU architecture are the introduction of NVIDIA's exclusive new high-speed bus nvlink, dedicated to high-speed interconnection of GPUs and CPUs, and the ability of the GPU to access system memory at up to 5 times times the traditional bandwidth of PCIe GB/s Hbm2,tesla P100 is the world's first GPU with HBM2 memory, with the fastest and highest capacity available, and has significantly improved the unified memory of the programming model to access all CPU and GPU memory in the system with a single unified virtual address. It greatly simplifies the portability of the program and the ability of data throughput.

"Microsoft is developing a super deep neural network with more than 1,000 layers," said Huang Xue Dong, chief voice scientist at research. NVIDIA Tesla P100 's amazing performance will enable Microsoft CNTK to accelerate the breakthrough in AI. ”

800,000 of AI servers, is it worth it?

The NVIDIA DGX-1 is priced at $12,900, or about 800,000 yuan. So, is the price really worth it?

NVIDIA DGX-1 offers 8 Tesla P100 accelerators, 16GB memory per GPU, 7TB SSD DL cache and more, with throughput equivalent to 250 E5 dual X86 servers. Then, according to 20,000 yuan a E5 server simple estimate, 250 units that is 5 million of the cost, which also does not include room, network, energy and other additional costs. The DGX-1 uses a 3U rack-mount chassis that can be used alone or integrated into a cluster, and is clearly more cost-effective with DGX-1 clusters.

In terms of integrated software, NVIDIA DGX-1 offers a complete set of optimized deep-learning software that is out-of-the-box. On the nvidia developer site Developer.nvidia.com, there is a deep learning in depth learning area, which provides the Deepin Learning SDK Development Kit, NVIDIA digits image classification and recognition software, Learning software, such as the open source framework, provides a full range of software support for deep learning and is available for download and use.

The Deep Learning SDK Development Kit contains powerful tools and class libraries for designing, developing, and deploying GPU-optimized in-depth learning applications. The class libraries include deep learning Foundation CUDNN, linear algebra, sparse matrices, multi-GPU communication, and a comprehensive cuda c\c++ development environment. The NVIDIA digits deep Learning Management scheduling platform provides pre-set optimization algorithms including Lenet, AlexNet, googlenet, and so on for image and video class data classification and recognition. In addition, Nvidia regularly updates the developer's website to provide developers with more optimization algorithms-if the GPU is already a part of the deep learning domain, the nvidia DGX-1 for AI machine learning will allow more businesses to move toward AI at a faster pace.

Commercial AI software based on the NVIDIA GPU also has a big advantage in the universality of the GPU: GeForce for PCs, Tesla for Cloud and supercomputers, Jetson for robots and drones, and all nvidia such as drive PX for cars The same architecture is shared by the GPU.

Baidu, Google, Facebook and Microsoft, the first companies to apply Nvidia GPUs to deep learning, have surged nearly 35 times to more than 3,400 over the past two years with Nvidia in deep learning, involving healthcare, life sciences, energy, financial services, automobiles, Manufacturing and entertainment industries.

Given the obvious advantages of Nvidia DGX-1 in terms of hardware, software and integration services, the 800,000 price is not high. NVIDIA DGX-1 is clearly able to dramatically increase the learning and training time of AI models, speeding up the processing of unstructured data from various types of images, videos, and voice on the internet of things, such as images of industrial production lines, medical video, road traffic images and video analysis, and so on, to help companies quickly, Benefit from AI algorithms as early as possible.

As one of the most important strategic partners of Nvidia DGX-1 in China, the world's leading supplier of surveillance products, China Safe City Solution provides the first single customer of Nvidia DGX-1, The latter will use DGX-1 for deep learning supercomputer projects in video surveillance.

NVIDIA DGX-1 is officially listed in July this year, and DGX-1 's listing is expected to activate the massive commercial of AI. For an enterprise, it is necessary to start thinking about AI strategies while investing in IoT projects. Before the big industry trend comes, only one step ahead can take the first step. (Wen/Ningchuang, "the era of Cloud Technology", No.: Cloudtechtime)

This article is from the "Cloud Technology Age" blog, please be sure to keep this source http://cloudtechtime.blog.51cto.com/10784015/1836665

Giants turn to the internet of things, the next stop in the internet of Things is AI