How should the semiconductor industry find gold in big data
Abstract: the application of big data analysis in the semiconductor industry is still in the middle stage (development period). On the one hand, in order to improve the performance, efficiency and reliability of chips, the amount of data in the semiconductor industry is increasing, and the task of data analysis is increasing; On the other hand, the semiconductor industry is also developing various technologies to fully tap the use of data
the amount of data generated by various devices keeps growing rapidly. How to use the astronomical data generated every day has become a difficult problem for the science and technology industry
according to Cisco data, the global interconnection throughput was 1.2zb in 2016 (1zb is equivalent to 1billion TB or 1trillion GB), which is expected to increase to 3.3zb per year by 2021. In 2016, the daily data flow increased by 32% year-on-year, while the 1-hour data throughput with the highest flow increased by 51% year-on-year
when the above statistics are not complete, in fact, no one knows how much data is generated worldwide every day, because not all equipment data will be connected to the Internet
the amount of data itself is of little significance, and how to quantify the value of data is the key, but there is no effective method on how to apply data to realize its value at present
if you want to mine the real value of data, you need to filter Tianliang digital and analog data, and fully consider the application scenarios. This is like looking for gold in sand, most of which may fail. However, with the improvement of computing power and the immaturity of large-scale parallel computing tools, data filtering analysis - that is, finding better methods to apply data - has been able to create a business model with promising market prospects
"many people in the industry pointed out that different data application methods are related to their devices and business models, and have an impact on their business, market and business models," said AART de geus, chairman and co CEO of Synopsys. "If you can find a shortcut, improve efficiency, or a new business model, it will have a great impact." This also means the possibility of high profits. "You will see that all people engaged in data processing are listening carefully to decode the future needs of the market, or judge the current market needs by themselves," de geus continued. "Or further, they are ready to be in the data path, so as to be closest to the center of data commercialization."
this is the reason for the crazy influx of capital. From data mining to cloud services, from machine learning to industrial IOT, every data application scenario is in fierce battle
"Whoever has data and has the ability to analyze and process data can earn all the money," said Wally Rhines, President and CEO of Siemens mentor business unit
at present, it is hard to say that data application is a winner take all game, but there are indeed many technology giants in this field, such as Amazon, Google, Microsoft, Facebook, IBM and other big guys who can output by breaking their fingers
"the collected IOT data contains a large amount of performance, behavior and application data of the device," said Christophe begue, sales director of IBM Americas. "Next, we will throw the collected data to Watson (IBM artificial intelligence platform) for analysis."
the big problem now is how to realize these data and who is willing to pay for the data. To realize data, we must first do the following: first, companies in the industry should really understand the value of data; Secondly, companies should be able to quickly respond to data changes. As long as they are one hundredth of a second faster than others, securities companies can make profits from this. But now large companies usually respond to data changes in a few days or even weeks; Third, the price of liquidation data should be competitive and not fluctuate too much
ibm is preparing to commercialize the data of the global supply chain. It can be set to 500X (the maximum measurable oscilloscope input voltage is 500 times). "Supply chain data is divided into two layers," begue said, "The first level is retail and fast moving consumer goods (CPG) And other data, which may affect the sales of food and beverage. You can collect relevant information such as weather, traffic or sports events in a nearby store and track it through traffic patterns. We use the Metro pulse platform for data analysis, which covers 500 data elements. Users can either buy data for in-depth learning or machine learning analysis, or entrust IBM to do all the analysis. The second level is that we are introducing the concept of market: supplier risk. IBM takes into account many factors such as weather and political changes, analyzes the security of the supply chain from the data, and improves the security of the supply chain according to the analysis results. If you notice that 15 factors are at risk, you will closely monitor these 15 factors. "
ibm's service is not value analysis of existing data, but also gives suggestions and insight into the future. "We collect public and semi public data, some of which are only used internally by IBM, and we build prediction models. Of course, we also realize that there is still a gap between planning and response, and the concept of 'decision room' helps to narrow the gap between planning and action."
not only the externally collected data is useful, but also the internally generated data is very valuable in industrial production. In fact, the whole concept of intelligent manufacturing (industry 4.0 in Germany, also known as industrial interconnection) is how to make good use of internal data
"in a word, industrial IOT is to improve production efficiency," said David Park, vice president of optimal+ marketing. "Now these companies prefer process analysis and inventory free production, but what they really need is predictive analysis. Predictive analysis can benefit factories, but brands benefit the most. Brands and factories are not necessarily the same thing."
the risk is that the data are not always correct. Decisions based on incorrect data will make the results difficult to predict
"if the data is OK, the yield can be increased by 2% to 3%, which is a very significant improvement," Park said, "All time period data of any component that passes the inspection in the supply chain will be collected. When you get some scratched wafers, you can find out which link the wafer is scratched according to the data, and you can also check the aging process of components on site. If the car is equipped with predictive maintenance services, you can see the relevant data of the car on the road. The financial industry will also benefit if you have hundreds of thousands of hair The relationship between invoices cannot be clarified manually. "
this kind of data analysis is particularly important for complex supply chains. Semiconductor manufacturing itself is advanced in data analysis and application, but the application of data in the whole semiconductor supply chain can not reach the level of manufacturing links
"effective use of data is a major theme of the Intelligent Manufacturing Advisory Committee (subordinate to semi)," said Tom salmon, vice president of semi collaborative technology platform, "Obtaining data is very important, but now the problem is not that we have insufficient data, but that the data utilization rate is only about 10%. The real challenge is what kind of questions we should ask and how to apply the data to manufacturing. Therefore, there may be reliability problems, but there will be no process problems."
find the key data, and use the machine to extrapolate the data rule under the preset parameters, which is the basis of machine learning. This method has been applied in the field of auto driving. Machine learning system will assist and eventually replace human beings to drive cars. Therefore, China has the development potential in this kind of high value-added materials. When making decisions, auto drive system needs to give a variety of forecasts according to the driving scene
in semiconductor design and manufacturing, machine learning will also be used to improve quality, reliability and yield
"using the appropriate proportion to extract data for analysis can be applied to future design," said Mike gianfagna, vice president of marketing at esilicon, "How to apply machine learning algorithms to new fields is the key. In the past seven years, we have accumulated a lot of experience in this field, and we know how to mine and develop the value of data. When you have a large amount of data, how to extract and analyze these data? If the proportion of extracted data is too high, you will get lost in a large amount of data, and if the proportion of extracted data is too low, you may not draw a conclusion."Gianfagna said that on the basis of reducing risk and increasing efficiency, realizing data realization is the goal of machine learning. "To do this, you need to look at big data analysis from an overall perspective."
compared with many big data analysis services provided by big cloud service providers, the semiconductor design and testing industry produces a small amount of data, but the semiconductor design and testing data may be more complex
"the main task at present is to collect data," said George zafiropoulos, marketing director of Ni solutions, "The goal of the next stage is to give improvement methods through data analysis. Can you find the value in the data without deliberately looking for it? What you are looking for is the trend and correlation of the data, and machine learning can be applied to any link. If the software prompts that the output of the production line this Thursday is low, why is the output low? Or the impact of specific temperature and specific voltage on product performance. (these can be guided by data analysis.) ”
zafiropoulos pointed out that better chip design can be the goal. "As engineers, we formulate rules around design, but if you want to cover all aspects, the efficiency will not be high. The overall demand for plastic machinery in China's domestic market will grow steadily. If protection rules can be reduced on the basis of ensuring reliability and performance, it will be of great value. A lot of big data analysis is aimed at multiple data collection points. A city may have 10000 sensors that generate a large amount of data every day, and Asia Masson's orders are countless. Semiconductor data is obviously more than personal data, but it is far from the magnitude of Amazon trading data. "
however, the system data may be several orders of magnitude higher than the design data, especially when multi physical layer simulation is involved. "We believe that 7Nm will be the first node to introduce machine learning and big data analysis. The amount of data will expand and the processing speed needs to be increased." John Lee, general manager and vice president of ANSYS, said, "You need to do synchronous thermal analysis. Thermal effect affects the reliability of the system, but if the amount of data increases to a level that current technology cannot solve, then new methods need to be introduced, so we need big data technology. The latest GPU has 21billion transistors and can be used in cars, but chips of this size have a huge heat generation, (if the heat dissipation design is not good) When heated, it will increase the pressure on the circuit board and may cause the board to bend, but you know, the service life of automotive chips is up to ten years. "
the application of big data analysis in the semiconductor industry is still in the middle stage (development period). On the one hand, in order to improve the performance, efficiency and reliability of chips, the amount of data in the semiconductor industry is increasing, and the task of data analysis is increasing; On the other hand, the semiconductor industry is also developing various technologies to fully mine data