Happy Holden Previews His AltiumLive Frankfurt Keynote
I recently sat down for an interview with Happy Holden, who is slated to give a keynote speech at this year’s AltiumLive event in Frankfurt, Germany. Happy gives a preview of his presentation, which is focused on smart factories and automation, and why artificial intelligence might improve PCB design and fabrication in the future.
Andy Shaughnessy: Happy, why don’t you start with a preview of your keynote presentation.
Happy Holden: There are numerous keynotes starting on Tuesday morning, and my keynote closes the event on Wednesday. Afterward, they have a raffle, where they will give away lots of great items for those that stick around to the end. With all of the high-speed trains in Europe, especially in Frankfurt, it’s easy to travel to this event at the Frankfurt airport, where the conference is being held. The train station is right below the airport.
I’ve been to AltiumLive the past few times, so it’s hard to think of a keynote topic because they’ve had so many good speakers, and they don’t like to repeat themselves. I’m not particularly an expert in design nor signal integrity, which the other four keynote speakers will address, such as Rick Hartley, who will discuss differential pairs. Instead, I will focus on elements of the future, and I picked an example of a smart factory—namely, GreenSource Fabrication in New Hampshire.
I have photos and videos to share from our visits. You’ve heard about it for the past few five years in Germany, but there hasn’t necessarily been a lot of visible progress. If the people weren’t aware of how we’ve covered the facility and the fact that it has no production workers, they wouldn’t know that it’s 100% automated. I give people a quick overview of the pilot plant, built in 2015, that was built relatively rapidly with many advances in technology that we haven’t seen before. It’s the first time that a factory has been put together with no production workers.
Once the robotic warehouse issues the material, the material never stops moving until it hits shipping about 208 minutes later, plus or minus a few lamination processes. If there are buildup layers, then it circles for a couple more hours before it’s shipped. Before that, Whelen Engineering, an automotive and aircraft OEM for electronic products, bought all of their print circuit boards from China. Their new automated factory, which didn’t cost them that much, had less than a three-year payback period, which means 50–60% return on investment.
Whelen was paying one-half to one-third of the price of the PCBs that they bought from China and receiving them in two days instead of 12 weeks with nearly perfect yields. This created a whole new paradigm for them in terms of competitiveness. So, instead of being a captive, in 2018, they built a new, state-of-the-art facility for the rest of the Western hemisphere, as a merchant.
They have imaging capability down to five-micron lines and spaces, and layer counts up to 36 layers, with every piece of material being electrically tested and categorized even before it’s laminated up. They have a lot of new imaging software—the part of the Industry 4.0 methodology that people talk about—and they have started to execute that. The ability to characterize every trace on a board three-dimensionally and characterize every piece of material out to 20 gigahertz allows them to do their controlled impedance of ±2% and kill a core if it’s not going to lead to that ±1–2% capability on the final impedance.
GreenSource is not talking about some of the innovative things that are leading them to build four or five buildup layers per side and still have 99%+ yields because nobody in the world has that kind of capability. Every prepreg and piece of foil and rigid has its own barcode space in the database for a characterization. The data storage must be unimaginable to be doing that many measurements constantly during the manufacturing process.
Shaughnessy: And that’s new for North America.
Holden: That’s a new thing for the world.
Shaughnessy: We see some of this smart factory stuff in Asia, but it doesn’t seem like we see it here as much. Isn’t the smart factory idea strong in Germany?
Holden: The Germans came up with the smart factory concept five years ago, and there have been investments in Germany in the smart factory, especially in assembly and the bare board. But they added capacity and capability instead of new people, but the people that were there remained. At GreenSource, there are around one dozen technicians who watch the process 24/7; everything else is done robotically and hands-off, as well as 100% recycling of all of their chemicals and water. Their costs are much lower than anybody else without labor and yield loss. That doesn’t necessarily mean it’s going to show up in their prices because they can be competitive on price, but they are likely to offer quicker turnaround time or nearly perfect yields and reliability.
If you don’t have people reading blueprints and setting dials on machines, then you can’t build the PCB unless you have all of the recipes digitized in the computer to download; Asia is not doing this. GreenSource must have a digital recipe for everything, which is a new challenge. Because of that, you have IPC working on IPC-2581, which is their new digital design data format.
That’s why the second part of my keynote goes into the IPC-2581 digital product model exchange, DPMX, which merges IPC-2581 with the CFX assembly data flow. They still have work to do to have all of the digital recipes you need to fabricate the bare board, but they’re coming along and hope to have that soon.
That leads to the last and third part of the keynote, which asks, “What do we do in the interim?” I go back to many years ago to Racal-Redac’s design tool, Visula, which had some interesting capabilities for the designer. They had these widgets on the screen that were based on the rules that you had set up for your company, or rules in the software that allowed for optimization for signal integrity. As you were placing, routing, and tuning, these little meters would be fine. But if you did something that was counter to good signal integrity, the meter would jump into the yellow or red; then, you could hit undo, move it back into the green, and choose a different technique.
I’ve always found the ability to take your company’s best practice rules on signal or power integrity and have them in the software, looking over your shoulder, useful for any designer. You had instantaneous feedback on things that may have been poor or questionable. You do that all the time now, but nobody ever sees the feedback until it’s time to build the prototypes. The fabricator comes back and tells you, “We had a hard time. If you had done this differently, it would have been cheaper and simpler. Unfortunately, the schedule doesn’t allow us to go back and redo the design.”
I finish with some research that we did at Hewlett-Packard Laboratories 25 years ago, focusing on DFM. We had a popular manual for designers that answered most of their questions. They wanted that manual automated, and the researchers at HP Laboratories created a self-learning AI. They had my group feed it 100 different HP designs, and this is the first time I ever saw of these deep-learning tools they talk about in AI. “Deep learning” wasn’t a jargon back then; they called it an “expert system that was self-learning,” but now it’s referred to as “deep learning.”
At first, we had a hard time figuring out how to make the AI learn, but we started with the simplest of PCBs—double-sided and low-frequency—and moved it through various types of analog and digital designs; we increased to higher speeds and more parts with more complexity and higher density. As we fed it more options, it learned, and we would have it design the board from scratch. We’d go back and compare how it had designed the product compared to how humans who designed it. After the deep learning, it designed superior PCBs in two to three hours compared to 2–12 weeks, and sometimes 16 weeks, to design.
However, it could never pick the best design. It would do 250,000 designs and give us the 10 best, but it still took an experienced designer to start the system rolling and pick out which of the 10 were the best or most marketable. There’s no way it could do 100% of every goal that we gave it; there was always a tradeoff involved.
In the future, there will be demand for more and more designs, but we may not necessarily have more designers. For the complexity of them, the costs would be high. And people look offshore to get their PCB designs, but what we don’t want is to outsource printed circuit design to China or India because if you lose control of design, you’ve lost the whole intellectual property of the electronic device.
Shaughnessy: What are some of the important things that you want the attendees to take away from your keynote?
Holden: The first thing is that smart factories with automation are coming soon. It’s not people waving their arms about an idea; there are now examples, and the people in North America who built them plan on building more in North America and Europe. The people who compete against smart factories will also have equally smart factories, which means we’re going to have higher-quality PCBs capable of meeting the geometries and the impedances and other characteristics that come with 5G and more specific electronics, so that’s good news.
The second thing is that, to run a smart factory, we have to have all digital data rather than blueprints and 70-page documents on manufacturing specifications. Third, we’ll probably get a lot of this done through AI since Intel and Nvidia now have specific AI chips. Their whole architecture of the computer is built around deep learning and AI. I’m hoping that the EDA tool vendors start taking advantage of these new chips, but maybe they’re finding engineers are in short supply, and they’re being gobbled up by everybody else in the AI arena.
Shaughnessy: If you mention “AI” to a room full of designers, some of them will get antsy and say, “We’re the AI!”
Holden: Twenty-five years ago, the AI we had was better than any printed circuit designer in the world in terms of meeting the objectives, but unfortunately, it didn’t know how to cross the last “t” and dot the last “i.” It would come up with the 10 top designs out of 500,000 designs but was not be able to tell you which one you should use in your product; that decision still required a human. But the AI isn’t there to replace them; it’s meant to increase their productivity and help them juggle all of the performance balls. AI will help manage some of those balls so that you can focus on the few that are the most important; let the machine do the rest.
Shaughnessy: It sounds like it’s going to be a pretty cool event. I’m looking forward to it.
Holden: I’m also doing a full-day HDI design workshop on Monday for Altium’s University Day, along with Rick Hartley and other speakers who are doing full-day workshops.
Shaughnessy: Thanks for speaking with us, Happy.
Holden: Thank you.