AI • IOT • NETWORK EDGE

The Industrial Revolution changed manufacturing. The rise of computers caused another upheaval of systems and processes. Now manufacturing is being revolutionized again by digital transformation and the advent of AI. This new frontier was on display in Germany this April at the Hannover Messe conference (HMI), one of the first post-Covid gatherings to not just talk about all the new possibilities available in the industrial environment but to showcase them.

At HMI this year were Ricky Watts, Industrial Solutions Director at Intel, and digital-trend expert Teo Pham. They discuss what they saw in Hannover—including ChatGPT— what it means for the manufacturing space—yes, ChatGPT in the manufacturing space—and how even the most exciting tech innovations could be pretty pointless if they’re not simple for users to implement (Video 1).

Based on what you both saw at the event, where do you see manufacturing heading?

Ricky Watts: In terms of technology, I think there are three areas that excite me and, to some extent, concern me a little. Some of the larger companies in this space are already using this technology to make manufacturing more efficient. To be honest with you, I was really surprised to see how much it was out there, and how advanced it was.

Another thing is this 3D reality—omniverse, metaverse, things like that—how immersive technology is going to be used in the future. Can I design and build a factory to change outcomes in a 3D virtual reality? And then use ChatGPT and AI to create digital twins that can create physical realities in manufacturing as well?

The last thing I noticed was a lot of robotics at the show. Robotics is everywhere in manufacturing, for good reason, of course—for the logistics and the repetitive tasks we often see in manufacturing. But one thing that was particularly interesting to me was robotsbuilding other robots to drive outcomes. One robot is given a task, and then another builds that first robot to drive the outcome of the task. The second robot uses AI to learn what it needs to do to send off a command to build or design a new tool for the first robot, therefore optimizing it.

Teo Pham: I was very surprised at the variety of topics and participants at this exhibition. Because you expect robots; you expect hardware manufacturers; you expect semiconductor manufacturers. But then you also had software companies, consultancies, and cloud-service providers. It just goes to show how varied this whole space is, and that it’s a lot more than just physical devices. This is how you create exciting new applications.

Ricky mentioned the metaverse. I saw companies like Siemens and Microsoft that were promoting things like the industrial metaverse, creating new technologies that make production more immersive, but also a lot cheaper in the sense of using digital twins to run these amazing simulations to really test things out in the digital space before needing to create them as a physical unit.

Which AI applications in particular got you really excited for the future of manufacturing?

Ricky Watts: That use of AI ChatGPT was one in particular. In the world of manufacturing, we have this thing called a manufacturing execution system—MES—or a programmable logic controller—PLC. It’s a device or an appliance that basically runs the manufacturing. PLCs have a language that operates and runs with them called 61131. One demo I saw was ChatGPT being used to build that code. Typically, a manufacturing engineer would write that object code, and it might typically take that person weeks or months to do it. ChatGPT was doing it in, I’m going to say, minutes or seconds.

I’m going to stress that it’s early days, what was being done in that demo. They were at pains to point out that there were some mistakes in the code, but it won’t be very long before the accuracy, and the ability to deploy that code directly to those machines, is really going to become relevant. Manufacturing is very much a structured environment, driven around a set of standards. But as we start to go into this new world, the possibility of being able to do that is really exciting.

Teo, where do you think the new opportunities are for manufacturers?

Teo Pham: Coming back to artificial intelligence, it just allows you to speed up processes, making them faster and cheaper. There’s always been a lot of talk about ChatGPT, but there are also AI tools that can generate images or blueprints or videos, websites or applications. I think the cost of generating 80% of a solution will go down to practically zero. But then obviously you will still need some very experienced people to get you from 80% to 100%. There will be some very fancy applications, like AI 3D modeling, for example, but I think even for fairly boring stuff like documentation or translation this will be very helpful because those tasks can be done within minutes.

What kind of processing power do you need to take advantage of these opportunities?

Ricky Watts: AI really relies on data. And once you’ve abstracted it, there’s the learning part and then the inference part. CPUs, GPUs, and also FPGAs are always involved.

A lot of the early use cases of AI in manufacturing have been visual ones: I put a camera into a manufacturing environment to analyze something, and then I train a model around the images I get. Let’s say I’ve got a production line of a bottle with a label on it. I’ve got a camera over those bottles, and I want to know if the labels are on correctly. So images are created around that, and then we would train models, generally in a GPU environment because it requires a lot of intensive processing.

Now I have something that knows what’s good, knows what’s bad. But in a manufacturing environment, I can’t keep learning all the time; it’s too difficult. So then comes the inference stage. I’m using the model, and I want to apply that. That’s where CPUs come into play, because it gets very tactical at that point where it’s very, very close to where the manufacturer is coming in.

The training is done where you’ve got a lot of compute power, typically in a cloud environment. The inference in most cases is done where the manufacturing is, at the edge. So you’ve got CPUs and GPUs, and both have an area of expertise. But what we’re starting to see from an Intel perspective is integrating them. You’ve seen it with some of our new technologies, particularly the latest Xeon^® chip, the Sapphire Rapids chip.

“#AI really relies on #data. And once you’ve abstracted it, there’s the learning part and then the inference part. CPUs, GPUs, and also FPGAs are always involved.” – Ricky Watts, @intel via @insightdottech

But now we’re starting to see compute platforms in these environments go from edge to cloud. In these environments there are two sets of data: the video I mentioned, but much more pervasive in manufacturing is time series data. Manufacturing uses what we would call fixed-function appliances—a machine, a robot, a conveyor belt—that are generating data that is not vision data. It could be heat, it could be pressure, it could be vibration, and so on. That type of data is optimized to run on CPUs at the edge as well. So you can do the training and the inference at the CPU, at the edge, where data integrity and data sovereignty are becoming very important.

On the CPU side I mentioned Sapphire Rapids. We’ve also got a new portfolio of GPUs coming out. It’s early days for Intel in that space. But we’re learning fast, and we’ve got more products coming out over the next few years. I think for us it’s going to be about integrating the hardware solutions, and then on top of that providing a uniform architecture for developers in the AI space to access those technologies, and we’ve built a number of toolkits and optimizations around that.

Irrespective of CPU, GPU, or FPGA, we optimize underneath; you tell us what the workloads are, and then we’ll deploy them into the right silicon platform at the edge and provide a uniform capability to take them to the cloud as well.

Can you expand on the benefits that manufacturers will gain from moving to the edge?

Ricky Watts: Manufacturing is a very competitive business—whether it’s physical items, like cars; or processes, like with chemical manufacturers. And the use of data in these environments can offer a very competitive advantage. It’s really about whether they can apply the technology in a business-driven outcome.

It’s very easy in our environment to forget that at the end of the day it’s not about the technology; it’s about the outcome. Go back to my bottle example. If I’m putting through a hundred thousand bottles a day, and let’s say 5% of them are inaccurate, I might be throwing away 5,000 bottles a day. That’s a sustainability issue; that’s a profitability issue. If I can reduce that failure rate to 1%, that has a massive impact on the performance of that factory.

What we need to do in the technology industry is make it easier for manufacturers to consume the technology. Manufacturers want to use it, but they’re not experts in AI, and they don’t always have data scientists. And we’ve got to make sure that it’s accessible for everybody in manufacturing, not just large-scale manufacturers that do have huge departments of engineers and data scientists. We are trying to give them all the easy button.

Teo Pham: When we talk about the implementation of AI, I think one of the decisions we have to make is about whether to do it with edge computing or cloud computing. Obviously there are some advantages to edge computing: It reduces latency. In terms of data privacy, we don’t have to send it to a cloud. On the other hand, we need to invest more in hardware, which can be costly and takes up space.

What are your thoughts on the “edge versus cloud” debate?

Ricky Watts: There are distinct advantages with both scenarios. Getting data into the cloud is very expensive because the volumes of data are massive. There are considerations around regulation, data sovereignty, privacy, security, etc. But there are a lot of advantages to doing training in the cloud and doing inference at the edge. Then as more and more powerful compute comes down to the edge, not only the training but the learning can be done there as well. So in my mind more processing will be going to the edge.

So, what’s coming next for this space?

Teo Pham: People say that we are witnessing the iPhone moment of artificial intelligence. Even before the iPhone came out in 2007, we had phones, but still the iPhone changed everything. Today we can’t even imagine a world without the iPhone, without smartphones, without mobile apps.

Similarly, AI has been around for 50 or even 60 years, but I think we are currently in this kind of virtuous cycle: We have lots and lots of data; we have the necessary compute; we have the models; and we have very easy-to-use interfaces like ChatGPT. So much progress is being made that maybe even in six to twelve months the whole space could be unrecognizable. We’re in for a pretty fun ride.

Ricky Watts: Ultimately manufacturing has to produce goods. So what I see is that manufacturers are focused on the new technologies, but they also need to make sure that the manufacturing environments they’ve got today are going to be there for the next few years.

Here at Intel, we’ll go on making sure that the manufacturers are generating the goods we need; and, if it’s energy, that the lights are kept on. We want to make sure that the transformations to come are smooth and integrated, and that there is as little disruption as possible.

Related Content

To learn more about smart manufacturing, listen to How Smart Factories are Revolutionizing the Industrial Space. For the latest innovations from Intel, follow them on Twitter and LinkedIn, and follow Teo on Twitter at @teoAI_.
 

This article was edited by Erin Noble, copy editor.