Published on 28-Jun-2013
"We chose IBM InfoSphere Streams as the only platform capable of dealing with our enormous data volumes." - David Sankus, Product Engineer, IBM Burlington
Big Data, Enabling Business Flexibility, Information Management Foundation
Established in Essex Junction, VT, in 1958, the IBM Burlington site has long been an important semiconductor fabrication plant—or “fab”. Burlington-built technologies enable Wi-Fi connections from smart phones, tablets, notebook computers and more.
With data generation growing fast, IBM Burlington needed to accelerate data collation and analysis to achieve near real-time disposition of anomalies in the semiconductor-wafer manufacturing process.
Applications based on IBM® InfoSphere® Streams software analyze hundreds of live feeds of data, enabling near-real-time anomaly detection and alerts, yield analytics, diagnostics and prediction.
Highlights critical alerts and reports for engineers’ immediate attention; helps minimize wafer loss and rework; gives visibility of quality issues before shipping; creates time-to-market advantages.
Established in Essex Junction, VT, in 1958, the IBM Burlington site has long been an important semiconductor fabrication plant—or “fab”. Burlington-built technologies enable Wi-Fi connections from smart phones, tablets, notebook computers and more. In addition to manufacturing high-performance analog and RF chips, Burlington hosts a major wafer and module test center for OEM chips for use in game consoles, such as the Microsoft Xbox 360 and the Sony Playstation 3, for IBM's ASIC [application-specific integrated circuits] Design System, and for the high-end server microprocessors that are manufactured in IBM’s fab in Fishkill, NY.
Testing is a vitally important step in the semiconductor fabrication process, and the aim is not just to avoid shipping faulty chips to customers. Profitability in the fab business depends on the yield—that is, the number of viable chips that can be produced from a single wafer. Manufacturers want to detect manufacturing process excursions as early as possible, so that they can avoid processing with those excursions and having to rework or discard entire batches of wafers at great cost. In addition, the ability to verify tests rapidly with predictable outcomes is critical to the efficiency of supply chain management, as the business needs to know as soon as possible if—due to defects—it needs to manufacture additional wafers to fulfill a particular customer order.
As chip designers continue to utilize transistors with shrinking dimensions by packing more of them into each processor, the volume of data generated during the testing of wafers is growing rapidly. The current generation of IBM POWER7+™ processors uses 32 nanometer (32nm) technology to pack 2.1 billion transistors including 80 MB of embedded Dynamic Random Access Memory (eDRAM) into each chip. The POWER7+ chip breaks down into 13 independently testable execution units, and there are 80 chips on each 300 mm wafer.
David Sankus, product engineer at IBM Burlington, comments, “Our ultimate goal in testing is to provide near-real-time analysis and reporting, with automated alerts to support the test and product engineers involved with the product. This requires us to sift through vast amounts of data. For the POWER7+ product, IBM needs to confirm that the 8.5 trillion transistors manufactured every day are working as expected.
“At 32nm, the data volumes generated during testing are tremendous. For a single product, we’re pulling out an average of 350 MB of data per hour, and that can rise to peaks of 1 GB per hour. Across the entire fab, with multiple products to test, the data rates could be 100 times greater.”
JJ Wu, distinguished engineer at IBM systems and technology group, comments, “With the increasing complexity of CMOS technology, which involves thousands of critical manufacturing processing steps to successfully fabricate a chip, those who can predict ultimate product behavior as early as possible in the processing sequence will have the greatest leverage to enable new product design tapeout first-time-right, to bring up manufacturing yields quickly, and hence, to win in the marketplace.”
Big data challenge
With the latest generation 22nm process IBM POWER8™ chip on the horizon, IBM Burlington knew that volumes of test data were set to grow by two orders of magnitude, further outstripping the capabilities of its existing tools for data capture and analysis. It chose to deploy IBM InfoSphere Streams software at the heart of a new end-to-end analytics platform to address this big data challenge.
“It was typically taking all morning to pull out the data and generate the hundreds of daily charts needed by the product engineers,” says Sankus. “The process was also very fragmented; we wanted to enable different test teams to share data and analysis more effectively. We chose IBM InfoSphere Streams as the only platform capable of dealing with our enormous data volumes.”
With that issue resolved, the next biggest challenge is to take the test results and integrate them with information from the fab to provide a complete picture of the end-to-end manufacturing process. In the near future, there will be eight distinct systems from the fab providing data—in total, 400 to 500 live feeds of raw data. IBM will be incorporating test data with the results of these live feeds of fab data, using multivariate analysis to produce yield prediction and diagnostics that support both the fab and the Supply Chain Management in near-real time. Multivariate analysis includes the use of Partial Least Squares (PLS), and other techniques, to highlight areas of the fab that impact the yield of the product.
Automated capture and alerts
IBM Burlington continues to use other semiconductor-industry analytical tools and reporting technologies alongside the Streams platform, which provides first-level triage and enables reliable, high-speed anomaly detection and mitigates the challenge of sifting through much larger and more complex data sets.
The Streams platform and toolkits allow automation of the capture of data from test systems, provide in-flight analytics, anomaly detection and alerts, and enable Burlington product engineers to produce yield diagnostics and predictions in a near-real-time setting. In a future phase, Burlington will use Streams applications to capture data directly from sensors, bypassing the database to deliver real-time processing capabilities.
“With applications based on Streams, we can carry out intermediate analysis and update charts as new data comes in, without having to spend four to six hours just gathering the data,” comments Sankus. “We have created Streams applications to produce rolling one-day and ten-day averages of all the different tests applied to each die. As each new wafer comes in, we can quickly see how it compares to those averages, and there are automatic alerts and escalations when specified limits are exceeded.”
With consistent 24/7 monitoring that runs deep analytics to correlate manufacturing process behavior and alerts from Streams applications, IBM Burlington is saving considerable time and effort in analyzing product test results. The solution enables fabrication to be modified early in the process if a systemic problem is identified, thereby reducing wafer rework or scrap and reducing turn-around-time required to meet fluctuating product demands.
“If you can find just one recurring problem and solve it fast, that can easily pay off the entire investment in Streams,” says Vijay Sankaran, associate partner, IBM Global Business Services®. “Each month, some fabs produce as much as USD100 million-worth of leading-edge products, and if you can positively impact just one percent, that’s over a million dollars saved.”
Identifying manufacturing problems earlier in the cycle improves order fulfillment that goes beyond yield management. Sankus cites an incident in which a faulty testing system was allowing bad parts to pass as good. “By the time we discovered the issue, multiple products in a seven-month long manufacturing cycle had to be recalled, delaying deliveries to customers. If the new solution based on Streams had been in place at that time, we would have identified the issue almost immediately and avoided a multi-million dollar impact.”
He adds, “The speed at which the supply chain moves puts pressure on the testing process, and there’s a danger that key triggers slip by unnoticed. With Streams we gain time: rather than being reactive and playing catch-up, we have event-driven applications that send out alerts automatically and tell us where to focus our attention. Today, as soon as the product engineers get into work, they can review their email alerts, then look at the appropriate summary charts to see what has been happening. In the testing process, what’s critical is the ability to quickly focus in on the five percent of data that really matters—that’s what Streams gives us.”
Transforming data into insight
IBM Streams software is highly flexible software designed to address the challenges of in-motion big data across all industries and all types of data. For any manufacturing environments that depend on the timely monitoring and analysis of metrics to maintain tight control over production quality and yield, Streams software provides the ability to zero in on the most critical information in vast, constantly changing sets of data.
Sankus comments: “The flexibility of Streams is a great help: We took the original application that we wrote for testing the 45nm POWER7® chips, and with small adjustments only, re-worked it for use on the 32nm POWER7+ and the 22nm POWER8 chips. We also adapted it for our application-specific (ASIC) products, which represent a very different set of requirements.”
While the processor products are treated as single parts managed by a large number of engineers, for ASIC products, a single engineer may be responsible for up to 100 part numbers. To identify systemic defects, the engineer must analyze vast quantities of data across many parts. IBM Burlington is using Streams to look across numerous part numbers, determine the yield for each type of array, normalize the data, and highlight anomalies. “ASIC is crying out for automation—this is an area where Streams has a really major impact,” says Sankus.
Faster and more productive
With Streams software, IBM Burlington has laid the foundations for an enterprise end-to-end analytics platform that will pull in live data from the fab and provide near-real-time analysis to enable manufacturing engineers to make decisions for handling possible manufacturing process excursions and project outlook, reducing supply chain uncertainties. With new technology and products introduced at an unprecedented rate, constantly validated models enable concurrent technology and design learning.
The result of the Streams implementation at IBM Burlington is to enable fast focus on anomalous results across hundreds of wafers and determine patterns of anomalies. Yield engineers then perform root-cause analysis to associate those patterns with events that took place in the fab during the processing of such wafers, and make changes to the manufacturing or design so that the next batch of wafers are of higher quality. It is a labor- and computation-intensive operation to perform such analysis on very large datasets. By finding out in advance the most probable locations of the proverbial needles in the haystack, the Streams implementation significantly reduces the amount of effort needed by yield engineers and the time to detect root causes of process excursions.
“This is just the starting point in our use of Streams, and people are already seeing significant benefits in the speed of access to the information they need,” concludes Sankus. “By enabling us to identify and analyze faults faster and more accurately, Streams is helping us to increase yield in our fabs, ultimately enhancing profitability.”
Time-to-market determines winners and losers in the marketplace of electronics products. With the increasing growth in the complexity of semiconductor technology, solutions that enable design tapeout first-time-right and reliable decision-making to respond to manufacturing process anomalies can provide significant competitive advantages. Streams has demonstrated the capability to integrate high volumes of data that are constantly generated from manufacturing processes and tests, provide deep analytics to characterize circuit sensitivities from process behavior in a low-latency environment, and provide reliable decision-making when manufacturing anomalies occur.
More accurate modelling of CMOS technology that is supported by enterprise infrastructure, such that the effect of process variabilities can be rapidly incorporated into design-enablement solutions, provides more predictable behavior for next product design tapeouts. Reducing the cycle time to validate and/or update designs with improved understanding of manufacturing process variabilities offers enormous competitive advantages in the semiconductor industry.
To get involved in the conversation, visit: www.ibmbigdatahub.com
Products and services used
IBM products and services that were used in this case study.
© Copyright IBM Corporation 2013 IBM Corporation Software Group Route 100 Somers, NY 10589 Produced in the United States June 2013 IBM, the IBM logo, ibm.com, Global Business Services, InfoSphere, POWER7, POWER7+, and POWER8 are trademarks of International Business Machines Corp., registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the web at “Copyright and trademark information” at ibm.com/legal/copytrade.shtml This document is current as of the initial date of publication and may be changed by IBM at any time. Not all offerings are available in every country in which IBM operates. THE INFORMATION IN THIS DOCUMENT IS PROVIDED “AS IS” WITHOUT ANY WARRANTY, EXPRESS OR IMPLIED, INCLUDING WITHOUT ANY WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND ANY WARRANTY OR CONDITION OF NON-INFRINGEMENT. IBM products are warranted according to the terms and conditions of the agreements under which they are provided. Actual available storage capacity may be reported for both uncompressed and compressed data and will vary and may be less than stated.