The Cook's Tour: AMD's only microprocessor fab is in Dresden, Germany. It is called Fab 30. The cleanroom is sized at 14K meters square and has a capacity of 20K WSM (Wafer Starts per Month). It was named fab of the year in 2001 by Semiconductor International. Today, many would say that it is just another 200mm fab. But dig deeper and you will find it's quite special. AMD has systematically matched and even exceeded the best with as little as a tenth the resources. Others have spent far more and most didn't come close. The results are pretty impressive. AMD is the only company to field a competitive commercial microprocessor. AMD was the first to have copper in production with yields above aluminum (something that many were still struggling to do four years later). It was the first to bring SOI to high volume microprocessor production. It has also kept up with the unmatched scaling needs of microprocessor transistors – on average, introducing a new transistor node every two to three months. The factory they've built here holds many of the secrets to AMD's competitiveness. So I'm going to delve deeper than the normal fab tour.

How do you compete head on with the world's largest microprocessor company? It starts with people. At AMD it's all about the people. It isn't just some fluffy mission statement on conference room posters; it's real to the core here. You can sense a lot of positive energy among the people at AMD when you're in Dresden. One way to tell is the degree to which quality discipline pervades the organization. A good reflection of this that I found inside AMD's fab was that tool boxes and PM spares kitting were a model of orderliness. It was easy to find poke yoke* techniques applied throughout. Instructions and procedures were well documented and readily available inside the fab. People have to understand and want to do this for it to be successful.

Morale is essential and it is running very high at AMD Saxony. One reason is that the facility has created 2000 direct jobs in a region that was very poor after the fall of the Berlin Wall. 40% of the people working at AMD Dresden have university degrees. More importantly, 30% of them were previously unemployed. Average blue collar worker wages for Dresden have risen 16.2% since 1997 and white collar salaries are up 23.6%. So there has been a significant benefit to the region as semiconductor production was introduced by Infineon first and then AMD. But don't think this is just something happening in Dresden. AMD has a long history of having highly charged-people.

Nevertheless, people are not everything. Thinking that your people are more motivated, or better, than your leading competitors can be a fatal mistake. It is what you do differently with them that counts. It's the break-away strategies they develop and the degree to which they are empowered to turn vision into execution.

AMD's strategy has always centered on being agile at the leading edge. There is a corporate culture here that centers on speed. This had been strengthened with Hector Ruiz coming to the helm because he is more hands-on when it comes to operations. It's not just time-to-market, it's speed everywhere. Within the factory, the focus is substantially different than the classical Management By Objective focus (MBO). One way that AMD gains its speed is in what it calls Management By Exception (MBE). They architect the objectives into the system so that they don't have to worry about these and then they focus on the exceptions. System is not just a word here. It is a real computer network that executes objectives automatically. At AMD, computers are used to automatically make the no-brainer decisions – that leaves time for the humans to focus on the exceptions. Another tool AMD has developed is something they call value stream mapping, which focuses on the value of the product as it progresses down the line. These tools are designed to quickly contextualize what's happening in the market with a manufacturing response. The result is much faster decision cycle times (DCTs) throughout the company.

One thing about having one fab is that you must do development in the same fab you're doing production. That means adding new processes and materials right in the middle of everything. A factory will be brought to its knees without superb controls. Just to give you an idea of how massive this job is, last year AMD routinely ran 2.4M MES transactions per day; 120K dispatch queries per day, and 15K carrier moves per day. And these numbers are growing constantly. What's amazing is that the number of masks has grown by over 50% and the fab is still chunking out the same 20K WSM it was designed to do almost ten years ago. So there has been a huge productivity benefit extracted. These benefits are the key reasons why AMD chose to develop its own intelligent factory automation. It is a system that far exceeds that of anything commercially available.

AMD's Automated Precision Manufacturing (APM) is one of the cornerstones of its ability to compete against more modern 300mm fabs. It does this not only by enabling greater efficiency, it also enables higher baseline yields, better binning, and faster yield learning, all of which drives product performance. APM is essentially a massive factory control software suite. Its IP portfolio is rich, with AMD having more than 250 granted and pending patents. To understand why it is so powerful, consider that it was architected with the following goals:

1. Enable flexibility and modularity of WIP, tool, and recipe to deal will many different products, rapid technology change, and aggressive process development in a production fab.
2. Enable fast lot movement and intelligent scheduling that shortens learning cycles in spite of growing masks by being able to set priorities, schedules, and tool loading dynamically.
3. Enable tight parameter control despite quick changes, process mix, and fast yield learning for new processes by building in control solutions that have a sturdy framework architecture.
4. Enable rapid yield learning and excursion suppression with fully integrated engineering data collection and analysis.
5. Automate decision making to a highly integrated level.
6. Architected to be easily modified and improved upon continuously.

These goals are still in place today – lighting the path for what are continuous improvements to APM. Change is constant at AMD and they average 65 software changes per week. It's not just the software that is changing, as they also make an average of 25 process flow changes per week and 1000 recipe changes per week. Being able to cope with this level of change proves the robustness of APM as well as highlighting the agility it brings to a company proud of being fleet of foot.

AMD's APM spans across and integrates product performance targeting, process control, production scheduling, yield management systems, and equipment performance optimization into one common IT system. What I found particularly impressive in its APC module (Advanced Process Control) is that it automates wafer level recipe control, fault detection, and e-diagnostics; along with integrating a next generation SPC (Statistical Process Control) system and fab-wide control system. 50% of the tools have fully automated APC control loops; 10% have wafer level recipe control; and 60% have controls to ensure that the equipment configuration is constant.

The depth and robustness of APM's SPC is superb as they have about 22K active SPC charts. These are rolled over on a continuous basis, as 100% of the tools have automated SPC data setup and collection. AMD adds about 30 new SPC charts every day. These are not just charts for a human to look at, systems watch and act on them as well. About half of the SPC charts can automatically put a lot on hold and a third can bring the tool off line (AMD calls this "automated lot-on-hold" and automated tool down”). Roughly 10% can do both. One of the more subtle benefits of this is that AMD has been able to extend out times between preventative maintenance (PM) and also been able to better integrate PM schedules with production schedules for greater efficiency.

There is also a YMS module (Yield Management Systems), which is an analysis tool that integrates data mining and allows chaining of engineering data from different tools. In other words, it allows engineers to bring together disparate data sets automatically into one common analysis platform. What I saw is a process engineer's dream. AMD's YMS has evolved from wafer level analysis to the die level and is now moving to a level of capability that is sub-die. While similar, YMS is fundamentally different from SPC in that SPC is used to control equipment, while YMS is used to diagnose problems and improve yield. YMS is much more of a machine-to-human interface. It serves the needs of three different groups: product line engineering, fab product engineering, and fab-wide contamination management. There is also a tiered structure to deal with yield issues at the module level, fab-wide, and at the product level. There are also two key monitoring points: process and equipment. Dealing with these widely disparate needs in a single system is very difficult, but the benefits are less duplication of effort and wider availability of data. AMD's YMS eliminates the data silos that form, too often, between groups.

The IPS module (Integrated Production Scheduling) has real time dispatching of wafers and active scheduling. In the latest version, the scheduling will be predictive and agent based. Each lot will be its own agent that carries with it all the information needed to finish, including multiple levels of priorities. Agent Based Scheduling (ABS) is formed around scheduling agents and processing agents. These agents are independent entities that serve and look out for their owner in the system. The agents can represent lots, machines, PMs (Preventive Maintenance schedules), resources, etc. The power comes from collective behavior of these agents. So if sales or engineering have a hot lot – even the ones that just have to go in today that didn't exist yesterday – it can simply be entered in and the system dynamically makes all changes needed to schedules and tool loading. Think of ABS like an anti-lock braking system for factories that keeps the company from skidding off the road when the market slams on the brakes. Better yet, in this case the four wheels act autonomously, while working towards a common goal. It's all real time, so managers don't have to wait to see what the impact on other lots will be. More importantly, because it is ABS it won't allow the new hot lot to override ones with higher priority. The agents for the latter will over-rule the agents for the hot lot automatically. These will still meet their deadlines with no human intervention. Now that's automated decision making and it's really cool.

Put all of this together and it brings even more power to AMD's competitive position. APM has enabled them to add more Internal Device Versions (IDVs), which in turn opened the door to massive differentiation. With APM, AMD has created an operating environment requiring that every single lot finish different from the device it started out as. They can change part numbers mid-process, anywhere in line. AMD has linked process recipe contexting with production planning so that they have modular routes that break up into reusable chunks. Modular routes allow for rapid implementation of new technology flows that are variations of prior flows. That is a fundamental change because it allows much better responsiveness to customer needs. The old formula was that when markets changed, a chip maker started more wafers and scrapped the ones in-process that weren't needed or finished them and sold them at a loss at fire sale prices. A wafer typically started and finished as the same part number. There were gate arrays, which allowed the last few interconnect mask levels to change, but these proved extremely wasteful of silicon. What AMD has done is drive the change to the process module level, as well as the functional block level. The important point is that they can change a wafer to a different part number mid-stream in the process. This makes them far more responsive to the market as well as being far more efficient, far less wasteful, and ultimately far more profitable.

How this is done goes back to AMD's value stream mapping focus and how they quickly contextualize what's happening in the market with a manufacturing response. As market opportunities emerge, the key is rapid alignment of AMD's production plan to the evolving needs. They first must decide what is needed, which is done with direction from Sales & Marketing via inputs from the OEM. They then prioritize, which is done with a combination of risk assessment, evaluation of market factors, and management input. This is then overlaid on their manufacturing position, which is an evaluation of what they have using manufacturing data. From this comes the decision point of what to build, which creates a new build plan that is inputted and out of which comes a new set of data that describes what can be built given the new plan, the resources at hand, and the Work In Progress (WIP). This is then turned over to a profit & loss planning process, out of which the final decision as to what they will commit is made. This entire decision cycle can go though a full revolution in 24 hours. To me, when I consider the companies that can't change the coffee brand in the cafeteria this fast, this is amazing. Any CEO will understand what this means to their business competitively and profit-wise.

As for the future, AMD's upcoming 300mm fab will integrate automated material movement with all these modules. The level of change going on here shows that APM is not something that will ever be an off-the-shelf suite of software. I've watched APM develop quietly from a vision I first heard about almost twenty years ago. A lot of credit has to be given to the development team, which executed on a vision established so long ago. At the same time, a lot of credit has to go to management for sticking with the vision for so long. I'm sure it hasn't been easy. But make no doubt about it: APM is a core strategic strength of AMD.

* Poke Yoke is a quality technique developed by Toyota to make the subtle sublime. With it, simple human error is virtually eliminated. The classic example is bolting wheels on. Toyota put a staging tray and a robot between the assembler and the open box of nuts. Every time a car rolled up, the robot placed five nuts on the staging tray next to the assembler. That way, if the operator was interrupted and forgot that some nuts had not been installed, it would be patently obvious because some leftover nuts would still be in the staging tray. Toyota did this across its assembly stations and quality soared.