Data Management Challenges In Heterogeneous Systems

Experts at the Table: Semiconductor Engineering sat down to discuss issues in smart manufacturing of chips, including data management, chiplets, and standards, with Mujtaba Hamid, general manager for product management for secure cloud environments at Microsoft; Vijaykishan Narayanan, vice president and general manager of India engineering and operations at proteanTecs; KT Moore, vice president of corporate marketing at Cadence; and Mohit Gupta, senior vice president and general manager of Alphawave Semi. What follows are excerpts of that discussion, which was held in front of a live audience at the recent Design Automation Conference. To view part one of this discussion, click here. Part two is here.

[L to R] Ed Sperling, Semiconductor Engineering; Vijay Narayanan, proteanTecs; Mohit Gupta, Alphawave Semi; KT Moore, Cadence; Mujtaba Hamid, Microsoft.

SE: One of the fundamental shifts in the industry is that a quarter or so of advanced chips are going into systems companies. The problem is we don’t get that data back because those are all being used internally, so the industry cannot learn from their mistakes. Moving forward, will more data actually be looped backward, or will it stay with the systems companies?

Narayanan: The issue is who owns the data, and that data will remain within the company. People in that company will learn from that data, and over time everyone will build on this knowledge. But it won’t be shared between one company and another.

Gupta: No, but that’s needed. When systems companies develop chips, they call it their secret sauce. But if you want to proliferate in the system you need to let it out, and that’s how you get more access into the system. If the fab guy has tuned a particular process recipe for your needs, they have to generalize it and distribute it to others for that to be economical for the fab. They may not provide that for the first few years to maintain the premium the initial guy paid, but eventually these have to be generalized and brought into the mainstream process.

Hamid: We shouldn’t expect that proprietary IP becomes available. The onus is on the design automation community, and the manufacturing equipment and manufacturing flow community, to determine what insights they need to improve the flow and the designers’ ability. That should not require proprietary design IP to be coming back.

Moore: When people ask us about our approach as to how we are implementing AI and machine learning, we want to make sure that the customer owns that learning. It’s not on us to keep it, but it is on us to enable it.

SE: Another issue is security. There are lots of components and materials and IP coming in from different parts of the supply chain, and equipment from multiple vendors. It’s no longer just about, ‘We’re going to make a chip.’ How do we secure all of this?

Moore: This is going to be tricky. It’s going to require collaboration of the whole ecosystem. And that’s going to be a challenge and an opportunity to see what we can do to make that work.

Gupta: When you look at the whole chiplet ecosystem, there are certain blocks we feel can be generalized and made into chiplets, or known good die, that can be brought into the market. The secret sauce is custom piece of silicon, and they can design and own the recipe around that. But there are generic components in any SoC — memory, interconnects, processors. You can always fragment it in a way that there are some general components, which you can leverage from the general market, and which will help everyone. That brings the cost of building your system down so you can focus on problems around your secret sauce. There is no need to worry about those generic functions. Chiplets will be able to leverage known ecosystems. It’s not there yet, but we believe it will get there.

Narayanan: We need something like a three-tier data management system, where with tier one everyone can access data and share it, and tier three is only for people in a company. But I don’t know when we’ll get there because data management is a real tough problem.

Hamid: We may need new approaches. Just looking at this from the hyperscale cloud perspective, which is huge, with complex hardware/software systems and things coming in from many vendors, how do we protect it? We can’t use the traditional approach. We have to use new approaches like a zero-trust framework. Multi-layered security is a defense in depth, like protecting a blast radius. When something happens, how do you protect it? It’s using concepts like enclaves, creating these cocoons so people can be working and they can’t go out. And then, it’s using all of the monitoring and AI abilities to be able to react almost instantaneously when some incident happens so you can catch it quickly and remediate. New approaches will be needed, and there are lessons to be learned. And then, new design enablement approaches are also coming. Many vendors, us included, were providing a service called Modeling and Simulation Workbench, where you have this secure chamber in the cloud. You can invite people to work in it, but everything’s monitored. We have data isolation, and all of those techniques. This was harder to replicate in an on-premise environment, but the new infrastructure allows the industry to do that.

SE: We’ve always designed chips for security. But what you’re talking about overlaying security, almost from the architectural level, allowing certain people access to data. Isn’t that a really difficult process to pull together?

Gupta: Yes, and it starts at multiple layers. You can look at it from an SoC level. These secure enclaves already are being implemented, where you want a certain function of the chip not being exposed when somebody is accessing the data. There are scenarios where a processor is sitting behind everything in the dark. Security has to be addressed. It is happening at a macro level in the cloud, which is where the data is residing, but it needs to be created from the ground up so you can connect all of this together and maintain the necessary security levels.

SE: One of the topics that seems to be coming up more and more is sustainability. Obviously, we need the chips to run at very low power, but we also need the factories to run very efficiently with a low carbon footprint. Where are we with that, and does that enter into smart manufacturing?

Gupta: People openly talk about this upfront whenever they come out with a new fab plan, whether that’s happening in Arizona or Texas or New York. The environment has to be a consideration. There’s power, water, and air involved, and all these elements need to be checked off in terms of your carbon footprint. These are highly automated fabs nowadays. Somebody was telling me that you can literally run a fab remotely without having anyone in there.

Moore: What we’re talking about requires a lot of compute power to analyze all this data, so data centers are going to continue to grow. And computer providers will continue to build faster, bigger computers with more CPUs, GPUs, and memory, drawing megawatts to gigawatts of power. With Industry 4.0, we’re going to continue to learn how we address sustainability. There are a lot of efforts from a design standpoint — software, EDA, and all of that — and we want to make sure we have the technology that enables our customers to meet whatever metrics they are setting as a target. It’s not just one size fits all. And it will require collaboration, because we can’t do this in a vacuum.

SE: How does generative AI play into smart manufacturing, and what kind of impact will that have?

Hamid: It’s very new, but for some applications it’s moving along quite rapidly. For example, we see the number of check-ins into GitHub, which is one of the largest — if not the largest — open-source code repository. Generative AI is approaching 50% of the number of lines of code checked into GitHub, which is staggering considering a year ago it wasn’t even there. That’s an example where it’s moving along really quickly. I believe it will go through the same cycle that all the other adoption cycles go through. For manufacturing, there will be scenarios for which generative AI is applicable, like doing some of the layout, and post-manufacturing. But we should not think that any one tool is a cure-all. You need to take a portfolio approach and see where this particular tool would fit in. It certainly will help the architect writing code, testing code, helping with floor-planning recipes, helping with layouts, reviewing logs, and providing feedback. All of those are heavy text-driven scenarios, and generative AI will do really well. We’re seeing two-hour-long Teams meetings being summarized into what are the key takeaways. For some applications, it’s a powerful tool.

Gupta: From an SoC design perspective, places where we should see generative AI are those places where a lot of manual effort is needed. SoC design and verification are very labor-intensive, and generative AI can be very helpful in ways that do not displace engineers and help speed up the overall design environments. I’ve heard people saying it can generate RTL code, but you can’t replace engineers because they are clever and innovative. Still, there are things it can do, like generate PowerPoints better than some marketing guys.

Moore: For assembling and building things in a manufacturing line, we’ve already automated a lot with robotics. But even in that, there are things that happen that we didn’t anticipate. So generative AI could be relevant if you apply it to mechanical robotics and manufacturing automation. It seems very possible.

View part two of the discussion: Need To Share Data Widens In IC Manufacturing

Related Reading
Fab And Field Data Transforming Manufacturing Processes
Data from on-chip monitors can help predict and prevent failures, as well as improve design, manufacturing, and testing processes.
Data Leakage Becoming Bigger Issue For Chipmakers
Increasing complexity, disaggregation, and continued feature shrinks add to problem; oversight is scant.