intel is the world's largest designer and manufacturer of sophisticated computing devices our story starts with over 110 000 employees in 46 countries and a record 72 billion in revenue in 2019 but our story is also about expanding the possibilities of technology for over 50 years to make the most amazing experiences possible not just for businesses and for society as a whole but for every person on earth how by driving the next wave of the information revolution by enabling telecommunications for individuals and businesses banking and finance commercial and recreational transportation manufacturing and service industries wholesale and retail businesses medicine education agriculture and more and recently intel technology has helped enable the explosion of cloud computing and big data analytics as well as the practical resurgence of artificial intelligence intel creates industry-leading silicon products that continue to get smaller more powerful and use less energy each year our complex manufacturing processes pack hundreds of millions of transistors onto an area smaller than the head of a pin we start with blank silicon wafers and after building sophisticated devices on them we saw the wafers into individual chips and put a number of them into a protective multi-chip package to realize our products designing these multi-chip products is just as complex with some chips containing as many as 10 billion transistors and 50 computing cores the results are some of the most complex products ever developed these products flow through at least six manufacturing stages at more than 200 internal and outsourced factories across 13 countries intel fulfills over 1 million orders a year from an international network of factories running 24x7 every day of the year in fact intel processes over a terabyte of complex supply chain data every day the intel story carries on as we use our expertise in design manufacturing and supply chain to dream even bigger powering the future of computing and communications while unlocking the full potential of advanced analytics and providing amazing experiences for every person on earth hello my name is naveen chinoy i'm the executive vice president and general manager for the data platforms group at intel corporation and my name is anne callahar i'm senior vice president and general manager of manufacturing operations here at intel the sophistication of intel's product design and manufacturing has grown as we seek to satisfy the world's insatiable appetite for computing 30 years ago our products contained one silicon die one processing core and a transistor count of just a few thousand today we are shipping products comprised of multiple dye over 50 cores and as many as 10 billion transistors this increase in complexity overwhelmed our previous set of decision support tools we have seen a similar rise in the complexity of our supply chain planning we need decision support tools that maximize product supply from our high volume factory network through the innovative application of advanced analytics this intel team built a powerful new set of decision support modules that has really helped ann and i make faster better and more integrated decisions the new set of integrated modules is applied from the started product architecture design to optimize the balance between performance and manufacturability through supply chain planning to optimize the use of our manufacturing facilities the results have been truly impressive delivering a total benefit of 25 billion dollars over the last 10 years both on the cost side and on the revenue side and since the new system is foundational and product independent uh financial and organizational benefits are going to last for many years to come thanks to the awards committee for giving intel the opportunity to present our success story now over to the intel team to describe the business and technical details of intel's innovative and beneficial application of advanced analytics hello my name is tony romero i'm a vice president in the technology and manufacturing group that anne co-leads and i serve as general manager of the customer fulfillment planning and logistics group as you heard from ann and naveen we design and manufacture a variety of semiconductor devices serving the world's computing ecosystem unlike many companies who limit their activities to only design or only manufacturing intel is an integrated device manufacturer that's one of our key competitive advantages today we'll show you part of how we utilize that advantage our goal is to satisfy customer demand by optimizing the use of our very expensive manufacturing network by integrating product architecture design and supply chain planning to achieve our goal we need to align supply and demand any misalignment results in unusable supply or risks unfulfilled demand and recovering unmet demand by generating more supply results in higher cost by applying advanced analytics we significantly improve our supply and demand alignment to deliver this capability our advanced analytics needs to manage four interrelated complexities the first is complexity is in our demand here's a typical set of server skus all these skus contain the same advanced compute core our architecture design addresses the need to serve many use cases requiring different features in a most cost efficient way for example cloud enterprise and network use cases need different numbers of cores operating speeds and levels of power consumption this server sku set provides a 56 core product for use in super computers as well as a four core product for small business servers and everything in between the red bracket indicates a subset we will use in a later example the second complexity is in our manufacturing process technology following moore's law to make smaller and smaller transistors on a regular cadence means that we can offer products with higher functionality it also means pushing the boundaries of physics resulting in more manufacturing steps and longer throughput times in our factories the third complexity is in our product architecture we've evolved our architecture from initially single die single core microprocessor solutions to now where many of our current products contain multiple compute dye each with tens of cores as well as memory chips graphics chips and input output chips each product can contain as many as 10 billion transistors we consider hundreds of thousands of potential product configurations to satisfy customer demand and assure manufacturability across hundreds of new products we introduce each year the fourth complexity is in planning our supply chain which consists of over 200 internal and outsourced factories spread over 13 countries considering ever changing demand routing 2. 5 million wafers of 400 unique types representing 4 600 unique skus across the factory network is a task that occupies hundreds of planners the supply chain planning steps in our overall process consider hundreds of thousands of routing possibilities per product to satisfy customer demand and optimize the use of our high investment manufacturing network which runs 24 hours a day 365 days per year considering the rising complexities in demand manufacturing design and supply chain over the past decade it's not surprising that previously non-integrated decision support tools developed for much simpler times proved to be inadequate our project fills the need for an integrated suite of tools and processes utilizing the methods of advanced analytics and operations research to realize the competitive advantage of being an integrated device manufacturer the result is a system that has achieved an integrated decision flow over our product life cycle from the beginning of product architecture design through the supply chain planning which drives execution within our manufacturing network throughout the product lifecycle the focus is on satisfying customer demand while optimizing the use of our very expensive manufacturing network with that brief introduction of our business problem i'll turn the flow over to evan to describe the initial steps in the decision flow hello my name is evan rash i'm a data scientist and principal engineer responsible for developing analytics that support our product architecture design process there are two main stages in product architecture design the first is architectural optimization where we focus on making sure that we've selected the best set of architectural options that meet our market requirements the second stage is performance optimization where we find the optimal balance of product performance and manufacturing costs i'll start by discussing the first stage architectural optimization so before we can choose the best architectural option we first have to understand what our customers need and this can be tricky because our customers have diverse requirements some care more about compute core and compute speeds while others care most about power requirements or other features to address these diverse requirements we segment the market into as many as 100 different schemes each with their own demand forecast if we tried to create a single superset design to cover all of those requirements the resulting chip would be too big and too expensive to manufacture on the other hand if we created 100 unique designs it would be too complex and too expensive to design and research this means that we need to balance the number and combination of design for example we could start with the theoretical superset design that has all of the features on it using the superset design we could then propose a variety of potential base designs that include some subset of the features and those are a1 b1 and c1 in this diagram then we could create derivative designs from those bases and tweak them so those are a2 through a5 b2 e3 and c2 in this diagram so how many designs and which selection of design should we end up building we can't afford to engineer all of them should we use a1 b1 and c1 or a1 a3 and c1 or some other number in combination of designs and this is a tricky question because there's a variety of complex considerations we need to take into account the first is the trade-off between engineering cost and manufacturing costs more design means a lower manufacturing cost because each design can be tailored to be smaller and have just the right set of features on it it also means a significantly higher r d cost support all those designs second consideration is the mix of demand at various options for example how much demand is there at high core count versus low core count the overall cost depends on the demand weighted product cost of our selected design the third consideration is how the device physics will vary between our options we use complex simulations to model the highly non-linear interactions between performance power consumption and the thermal properties of our proposed designs and these are just some of the considerations there are many others that we don't have time to talk about right now to address these challenges we developed an optimization model we start with the design options and there are typically 5 to 20 of these per product our manufacturing process forecasts our demand forecasts and our financial models then we feed the design options and the manufacturing process forecasts into our device physics simulations and we generate around a thousand manufactured parts we feed those into the mixed integer program to optimize the selection of designs in such a way that all the market requirements are met with the minimum manufacturing cost a single scenario typically runs in around five minutes and has a thousand binary variables fifteen thousand regular assignment variables 2 000 constraints and about 300 cuts users typically run around 5 to 10 rounds of analysis per product each with roughly 10 to 100 what-ifs with around five new products being analyzed every year the end result is a recommendation of which design options to pursue which is approved by the chief architect and the sales and marketing team as the product progresses further in its life cycle additional rounds of analysis may be required to assess changes in market conditions manufacturing forecasts or other engineering considerations this optimization model has yielded substantial cost savings in the example shown here the lifetime manufacturing cost of the product was reduced by 750 million dollars after applying the optimization model the initial engineering baseline that was generated with traditional techniques is shown above and you can see it consists of four different designs a 12 core 835 core a 44 core and a 60 core design the optimization model achieved the cost savings by counter intuitively making the smallest design larger increasing it from 12 to 21 cores and also by eliminating the 35 core design this results in a big shift of demand in the 14 to 20 core range being moved from the larger 35 core design to the smaller 21 core design the solve also lowered costs by giving the 44 core design additional medium core demand so that failures and cores in this design could be recovered this reduces the failure and scrap rates and lowers our cost so now that we've decided on the architectural options the next step is to find the optimal balance between the performance of our skus and their manufacturing cost we call this performance optimization all of our products are manufactured from a common forecasted material distribution so if you sample parts coming out of the factory you'll notice that they have a distribution of power performance thermal and other characteristics and we fulfill our demand by assigning to skus from this common distribution of parts so all skus are competing for the same material the performance of our skus is determined by the minimum requirements that we define for each student increasing the compute core the compute speed the cache size the power or other minimum requirements for our skus will increase the performance of that sku and also its selling price but that will also drive up the cost because less of the common distribution can be utilized so the question is how can we determine the minimum requirements per sku that give us the best performance versus cost trade-off and this is tricky because each product has around 100 skus and each of those have dozens of minimum requirements that need to be defined so we're searching a highly non-linear space that has over 10 to 100 possible solutions to address this challenge we implemented a genetic algorithm that iteratively searches the requirements space and then runs a linear program to optimize the sku assignment and minimize the cost the algorithm takes a baseline skew set the demand forecast the forecasted manufacturing process and weights for how important each sku is based on the sales potential of that sku these are then fed into the genetic algorithm which uses mutation and crossover operations as well as targeted heuristics to tweak the individual skew requirements these updated requirements are then fed into the device physics simulator to evaluate how much of the common distribution will be able to meet those requirements then we run the assignment linear program to optimally assign material to each skew and as the algorithm goes from generation to generation it explores thousands of possible solutions and it generates an efficient frontier tracking the best trade-offs between performance and cost by applying this algorithm we've realized substantial revenue improvement the example shown here demonstrates that the algorithm found 235 million dollars of incremental revenue by optimizing the sku performance targets now here the x-axis is performance and the y-axis is cost and each dot represents a candidate skew stack where one or more of the performance targets have changed the blue dots are the manually generated engineering baseline and as you can see the solver has highlighted the efficient frontier and has also recommended the green star solution this solution has substantial performance improvements for a very minor cost increase compared to the baseline resulting in 235 million dollars of increased driven this capability has also enabled a paradigm shift where senior management can make strategic decisions about where on the efficient frontier each product and each segment should lie management may choose to stay on the left side of the frontier for cost sensitive markets and shift more to the right of the frontier for performance-centric markets this was previously impossible due to the vast size and complexity of searching the space manual now that we've optimized the architecture and the performance of our products the next step is to optimize the manufacturing flow so i'm going to turn this over to john who will describe how we optimize the manufacturing process for our products hello my name is john heine and i'm a business architect in intel's global supply chain organization evan just explained how we've improved the way that we design our products once that design is determined intel faces the challenge of deciding how to optimally plan and manufacture those products in our worldwide manufacturing network this is the responsibility of our supply chain planning module in this section we will discuss the scope of supply chain planning the benefits of decomposing the planning problem and how we decompose the problem into three steps wafer optimization network capacity alignment and optimized routing plans the output of a supply chain plan is a schedule for every item location and route for each week over a one-year period due to the complexity of intel's products this is a very large scale problem with over 37 000 items 22 million routes and 200 internal and external manufacturing facilities intel's capital expenditures over the last 10 years total more than 100 billion dollars so optimizing the use of that investment is critical to our financial performance we need to generate a schedule that supports the demand and inventory target strategies is feasible within our manufacturing capacity and minimizes our total cost we also need to frequently update this schedule to align to changing business conditions each month we reset the plan for all products and each week we update the plan for key products based on business needs due to the size and complexity of the business problem we decompose it into the three steps shown previously which allow the planning teams to generate a solution that balances the competing objectives decomposing the problem into multiple steps along with advanced optimization techniques allows us to run a solve in just minutes instead of the hours it might take using conventional solutions it makes the results easier to explain it allows different planning steps to happen at different frequencies and it's used by hundreds of planners at intel not just a small team of operations research experts these techniques enable an iterative scenario analysis approach that allows us to combine business knowledge about intel's dynamic and rapidly changing business with high performing optimization capabilities that show the impact of different business decisions each time we start a new planning cycle our key business decision makers provide a set of inventory strategies and product priorities however due to the complexities described previously they struggle to predict where we will face constraints and problems in the supply chain how severe they will be and how they should adapt and adjust their strategy to deal with those problems in our iterative process our planners run the solve engine analyze the results work with their key decision makers to get feedback on the business trade-offs and scenarios they want to evaluate create additional scenarios to explore the range of potential trade-off decisions that could be made and go through this loop multiple times within the course of the day before aligning on the one scenario that they want to execute and publish to the rest of the supply chain as our operating plan with this approach the strategic decisions are made by the business operations teams but the execution is done by the optimization program the operations teams can explore many possible scenarios while the optimization engine shows them the best possible results they could achieve for each set of strategic decisions in the first step of our problem decomposition we focus on our wafer fabrication and sort facilities that consume the bulk of our capital investments we model the full product flow complexity to understand how wafer decisions impact the final finish good supply but we ignore capacity constraints at all other stages of the supply chain this model runs simultaneously for all products at intel it has 13 million decision variables 4 million constraints and can typically be solved by our engine in about six minutes because these decisions are so critical for intel success we run dozens of scenarios and carefully analyze how these decisions impact our capability to support customer demand and inventory strategies in the next step our wafer capacity allocation decisions and schedules are frozen and we know how much supply we will be generating on each wafer we use those schedules as a constraint when we generate a capacity request to all of our die prep assembly test and finish facilities both within intel and at our external manufacturing partners in this process each product family can be solved independently because the shared wafer capacity has already been allocated and capacity all other stages is unconstrained this breaks down one large problem into hundreds of smaller and easier to solve problems our internal and external factory partners then analyze the requested schedules and allocate their capacity to meet the demand and product priorities provided analyzing and allocating our internal factory capacity is a complex problem with over 7 million decision variables 7 million constraints that we can solve in about 4 minutes in previous steps we allocated the fab capacity and determined the supply available on each wafer then we generated a capacity request and received a response for all of the other stages now we can run a fully constrained solve to optimize our final factory routing schedules and provide detailed instructions to each factory on what they should build in every work week once again each product family is independent allowing the problem to be decomposed into hundreds of smaller pieces a typical large product family at intel has about two million variables 1. 6 million constraints and a solved time of less than two minutes the optimized routing plans are published to our enterprise systems and are used as the plan of record for customer order confirmation factory execution and for our cfo's financial projections the size and complexity of these models are very challenging especially considering the business requirements for rapid runs to enable many iterations and scenario analyses and a high number of objective trade-offs in a single optimization we have employed many innovative approaches and used our deep understanding of intel's business to exploit the structure of the problems we solve in the product design steps we integrated multiple techniques genetic algorithms device physics simulation and linear programming models into a single solution in the subsequent supply chain planning steps we're using a mixed integer program model extended with many advanced approaches like model decomposition sequential goal programming advanced cuts path-based strong formulation stage driven data models and constraint caching to effectively support our large user base we use knowledge-based systems and machine learning models to help explain the results and perform sensitivity analyses so how does the solution end up making a difference to intel's results that's what we'd like to share next this chart shows intel's revenue and the portion of that revenue that is managed by the capabilities we've described this project helped manage 93 percent of intel's revenue over the past 11 years totaling 568 billion dollars we've operated in three very different business environments over that time frame excess capacity versus demand prior to 2011 a mixed environment up until about 2016 and a highly capacity constrained environment up until today over the years operations research and advanced analytics implementations at intel have resulted in benefits in the range of five to fifteen percent typically the more complex the business problem the higher the benefit at intel we follow a well-defined process to ensure proprietary solutions are only used if there's a competitive advantage to be gained otherwise we use off-the-shelf solutions for product architecture design our business problems are unique and no off-the-shelf solution is available for supply chain planning we found off-the-shelf solutions to be lacking on several fronts primarily in their inability to effectively model our product and supply chain complexities including manufacturing constraints business rules and strategies this is an absolutely crucial criteria for our results to be executable also we found that off-the-shelf solutions cannot meet the performance expectations to support our iterative business process and time to market for new features does not meet the needs of our fast changing business in 2009 we replaced a legacy solution that used a greedy heuristic with our supply chain planning module we ran both solutions in parallel for a year the new solution showed that we could improve our results by six to fifteen percent with an average benefit of twelve percent since then the supply chain planning module has undergone major enhancements to effectively handle all of our business environments as well as our increased product and supply chain complexity these complexities have also rendered our legacy heuristic solution obsolete to quantify the benefits of this project we use a conservative six percent improvement assumption which is the smallest improvement that we observed over the one-year comparison sheila will now share with you the financial benefits that we realized from our idm analytics solution hello my name is sheila glaser i'm the corporate vice president chief financial officer and chief operations officer of intel's data platforms group i am a long time customer and beneficiary of our advanced analytics team i'd like to quantify the benefit our stakeholders and employees have enjoyed from the application of the advanced analytics to both our product design and to our supply chain first i'd like to cover the financial impact of the work of the team the conservative six percent improvement you have just heard uh described applies to both cost and revenue on the left-hand side of the plot in the years where intel had excess manufacturing capacity the six percent meant that we were able to manufacture six percent less wafers to meet demand resulting in decreased manufacturing costs on the right hand side in years where we were capacity constrained with the six percent improvement we could satisfy six percent more demand than we could have with other methods translating into increased volume and increased revenue in the middle years we oscillated between excess and constraint three percent was credited as cost savings and three percent as revenue uplift these valuations use average manufacturing cost and average revenue per wafer across the product line the most recent three years include the cost reduction and revenue contributions of products designed with our advanced analytics as they went into production the financial evaluation of the design modules is taken from internal awards given at the successful completion of design projects as computed by financial business and engineering personnel involved with the individual products the overall financial impact has been 25.
4 billion dollars so far the advanced analytics has also generated organizational benefit here are just a few examples of the internal wars in the upper left a divisional award recognition from even chenoy executive vice president of our data platforms group who you heard from in the opening video in the upper right a technology and manufacturing group excellence award from senior vice president ann kelleher who also was featured in our opening video the excellence award is the most prestigious recognition awarded by our manufacturing and operations group and must be approved by our executives e-level leaders in the lower right is an external recognition from gartner one of our industry analysts on this foil you can see comments from users and managers of users of our advanced analytics these include comments on the increased quality of the solutions generated the decreased time and effort required to generate solutions and the facility to solve important problems engineers and planners had never been able to consider before in the upper left a comment we have heard over and over during the development and production of our modules i started as a skeptic i became a believer shown here in the next foil is the environmental impact of the application of advanced analytics our manufacturing processes require a large volume of water which we purchase from local municipalities we clean the water to our high standards use it in factories and return most of it to the municipality much cleaner than we got it free of charge considering the same three capacity conditions as shown earlier along with our six percent improvement application of advanced analytics has reduced our water usage by 2. 4 billion gallons and eliminated 0.