Big Data Analytics dimension belong to a cognitive computing. Compared with large data, a wider range of cognitive computing, technology is more advanced.
IBM cognitive computing is a concept put forward that "cognitive computing" is through natural language to communicate with people and to continue learning, to help people achieve more systems, from hardware architecture is to algorithmic strategies, from program design to combined with a number of academic fields such as industry expertise, will enable people to better get more insight from vast amounts of complex data, in order to make more accurate decisions. IBM clearly defined as the cognitive computing - have large-scale study, based on objective reasoning and the ability to interact with the natural human system.
Cognitive Computing and Big Data Analytics What's the difference?
Big Data Analytics dimension belong to a cognitive computing. Compared with large data, a wider range of cognitive computing, technology is more advanced.
Cognitive Computing and Big Data analysis has similar technology, such as large amounts of data, machine learning (MachineLearning), industry models, large data analysis, more emphasis is to get insight, predicted by these insights. In addition, the traditional big data analytics will use the model or machine learning, but more by the experts.
For calculation cognitive, insight and is just one of the prediction. However, cognitive computing more emphasis on natural interaction between people and machines, these dimensions are not the traditional big data analytics emphasized.
In addition, a field of cognitive computing is growing rapidly as the depth of learning (DeepLearning), its learning with traditional methods, the more is to get such a model self-paced study based on large amounts of data without the need for a lot of human intervention, in terms of learning from this large data analysis and there are many different places.
The bottleneck of traditional computing resources CPU, GPU, ASIC, etc.
Over the past decade, AI went to the stage of rapid development. Depth study played a pillar in its development, despite strong predictive power of simulation, the depth of learning is also facing large computational problems. On the hardware level, GPU, ASIC, FPGA is to solve large computational programs.
In 2006, people still deal with the problem of machine learning with a serial processor, then Mutch and Lowe developed a tool FHLib (feature hierarchy library) for handling hierarchical model. For the CPU, the amount of programming it requires is relatively small and there may be benefits of mobility, but the characteristics of serial processing becomes its shortcomings in the depth of field of study, and this disadvantage is fatal. Today, according to 2006 years have passed, the development of integrated circuit over the past decade or follow Moore's Law, CPU performance has been greatly improved, however, it does not make depth study and research into the CPU again vision's. Albeit on a smaller set of data can have a certain CPU performance computing power, multi-core parallel processing so that it can, but this depth study is still not enough.
GPU: Although researchers into the line of sight, compared to the CPU, GPU core count greatly improved, which it has a more powerful parallel processing capabilities, it also has more powerful ability to control data flow and storage of data . Carried out on a difference between CPU and GPU processing object recognition in Chikkerur, final GPU processing speed of the CPU 3-10 times.
ASIC: application specific integrated circuit chip (ASIC) customized due to its characteristics, is a more efficient method than the GPU. But its customization may also determines its low mobility, once dedicated to design a good system to migrate to other systems is not possible. And its high cost, long production cycle, such that it is in the present study is not considered.
FPGA is the future of computing?
FPGA: FPGA and GPU made a trade-off in the ASIC, a good balance between processing speed and control. On the one hand, FPGA reconfigurable hardware is programmable, so it can be more powerful regulatory capacity compared to the GPU; on the other hand, the daily growth of the gate resources and memory bandwidth makes it more design space. More convenient, FPGA also eliminates the process flow sheet in the required ASIC solution. One disadvantage is that it requires the user to FPGA using a hardware description language can be programmed. However, there have been technology companies and research institutions to develop the language easier to use, such as Impulse Accelerated Technologies Inc. has developed a C-to-FPGA compiler to make FPGA more fit the user's, Yale's E-Lab is a Lua script development language. These tools are to a certain extent, to shorten the development time researchers to make research more simple.
Linux On Power + GPU + FPGA = cognitive computing for the future
Innovative Linux on power combined with FPGA, GPU computing model, this may be the best framework for the future of computing knowledge.
IBM announced in 2015 with the FPGA chip designer Xilinx "strategic cooperation for several years." The two companies will join forces, by OpenPOWER Foundation committed to finding better ways to deal with machine learning and network functions virtualization (NFV), gene, high performance computing and big data analysis applications.
IBM developers will build solution stacks for OpenStack, Docker and Spark, combined with the POWER-based servers, and have the Xilinx FPGA accelerator.
In addition, McCredie also announced that the company would POWER 8 chip with Nvidia Tesla K80 GPUs combined together, using NVIDIA's high-speed Internet NVLINK. Two OEM - Penguin Computing and E4 Engineering will design the system based on OpenPOWER market.
According to IBM internal testing showed that compared with Intel's E5-2699 V3-based server processor, the new Power Systems LC server can be less than half the cost of running the former newsletter Twitter analysis, page view displays and other data-intensive workloads for customers every dollar of costs provided ahead of the former 2.3 times the performance. With its efficient design, LC server can run more than 94% of social media Spark workload than the equivalent Intel processor-based servers in the same rack space.
Calculating for commercial and high-performance computing 2U Power Systems S822LC socket 2 with up to 20 nuclei, 1TB memory and 230GB / s memory bandwidth. S822LC also equipped for high-performance computing NVIDIA Tesla accelerated computing platform assembly Ultimate - two integrated NVIDIA Tesla K80 GPU accelerator. And similar configuration, compared to the x86-based server E5-2699 V3, two S822LC products provide leading to the former more than twice the performance of single-core, 40% higher than the former and more than twice the price at full memory configuration memory bandwidth.
Highly recommended reading articles
40 + annual salary of big data development [W] tutorial, all here!
Zero-based Big Data Quick Start Tutorial
web front-end development based tutorial
Basics tutorial to learn linux
Big Data engineers must understand the concept of the seven