Need To Share Data Widens In IC Manufacturing

Experts at the Table: Semiconductor Engineering sat down to discuss issues in smart manufacturing of chips, including data management and grounding, 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.

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

SE: Data needs to move all the way from design through manufacturing and into the field, and then back again to be really useful. Is this happening?

Narayanan: Yes, and I can give you an example. We had a case where the customer was seeing heat loss into the system level. We mapped it to a particular transistor parameter using our ML algorithms and cloud data. That goes back to the manufacturing person asking, ‘Can you just tweak the process?’ They saw improvement in the yield of 2% to 5%, which is a very good number. This is one example of feedback.

Gupta: If you talk specifically about the DFM or DFY, things are getting complex for sure. As we go to smaller geometries, things are not easier. We’re going to start counting electrons in a transistor as we go down below 1nm on a nanosheet. But manufacturing is also evolving. This is where the loop goes back into smart manufacturing. They are using these AI tools and engines indirectly to be able to predict what can go wrong? If you look at design rules, they’ve grown by a factor of 100 over three process nodes. That essentially means you can control the process at that level of detail. So there is some improvising, but we also pushing the innovation as far as possible. I have utmost trust in engineers to solve it.

Moore: It’s been a while since I’ve looked at manufacturing data. But all the data that came out of our fab that we had to analyze, from SPICE parameters to extracted data, was overwhelming for a team of engineers to look at. You’re doing it based on wall clock time, and you’re doing it serially. You might have one assumption or one takeaway from the data, and I might take away something else from the data. But we’re not looking at it holistically. With the advent of Industry 4.0, what makes it possible and exciting is that we have the compute power to address large datasets, whether it’s in the cloud or on-premise. And these algorithms that can evaluate the data, evaluate the learning from past designs, and push it forward are possible in a realistic amount of time.

Hamid: A third thing to add is design for trust. We’ve been doing work with the U.S. government on data that’s scattered throughout design and manufacturing to ensure integrity of the design. That will get important commercially, as well, particularly in critical industries where chips are going everywhere.

Gupta: Even today, you engage with various kinds of customers where security is the hot topic. Trust has taken center stage in terms of whatever you do, at whatever level. How is your data being kept secure, whether it’s hardware or software?

Narayanan: I agree with KT’s point of view. We need to move Industry 4.0 into the semiconductor domain. Cloud, AI, 5G, ML, IoT — all of these are very, very important.

SE: One of the problems with AI is understanding whether your data is good and your results are accurate. Can they be proven and repeated? Where are we with this, and how does it fit into manufacturing?

Hamid: We’re evolving. You have to treat the AI accuracy for machine learning, reinforcement learning, deep learning, and separate that from ChatGPT. We call it data grounding. It’s separating the two. Looking at the first one, more from the manufacturing standpoint, it all depends on the quality and availability of data. In many industries there is a lot of data available on which the models have been trained. A lot of progress has been made in the industry for synthetically generating data, especially for training self-driving car algorithms. There’s a lot of synthetic data generation going on, as well. For manufacturing, the usage has picked up a lot, and I’m seeing this across a wide swath of industry sectors. Semiconductor manufacturing actually is lagging other industries because of the concern around security and privacy of semiconductor manufacturing. As to the generative model, it’s still very early going. The concerns raised are more around data grounding. We have started to put up information on how we are grounding the data, and whatever data grounding is not available, making that clear to the user so they know how much to depend on the results.

SE: We’ve always looked at the data that comes out of the fab as golden, even if fabs didn’t share much of it in the past. But it becomes more difficult to set up a platform like that when you’ve got multiple chiplets, heterogeneous integration, as well as much smaller runs than in the past because you’ve got a lot more custom silicon. How does this affect smart manufacturing?

Moore: We need to be more specific when talk about smart manufacturing. There’s design for manufacturing when you talk about the silicon. There’s design for manufacturing when you talk about the board or chiplets. There’s design for manufacturing when you talk about the box or the casing that device is going into. It’s always been discrete, unconnected steps. You do your sign-off at the chip level and you’re done. Now it’s in a board, and somebody else has to look at those design for manufacturing rules. We talked about a data platform earlier. That’s what will enable us to cross these different silos or boundary conditions.

Gupta: You brought up the point about heterogeneous chips with chiplets. That’s still evolving. We have been a strong believer in leveraging chiplets to build systems. We’re getting to a place where now at least the chiplets can talk to each other. But there are still a lot of things that need to be clarified or sorted out by the community. If I pick up two known good dies, one from a fab in Taiwan and one from Korea, and put them in a system, will they work with each other in a solution that is golden from the get-go? There are challenges to be solved there. There are steps being taken by some consortiums. Protocols are being defined where things can talk to each other. But on the manufacturing/packaging/test side, there needs to be more of a push so all this data can be looked at together. ‘A’ needs to be able to talk to ‘B,’ which can talk to ‘C,’ and ‘C’ can also talk to ‘A.’

SE: We also tend to lump all of AI into one bucket, but it’s really more complicated. If you think about reinforcement learning and generative AI, those are very different from using training data for inferencing. What actually works in a manufacturing setting for smart manufacturing? Is it all the above? Is it for very specific purposes? Does each one have a role? Or is it just one or two of the possible sub-categories under AI?

Narayanan: It’s going to be a mix. It’s not just going to be AI. It could be machine learning, as well. AI alone may not be enough. You may do some kind of training of a machine learning model, which can put it in the right spot.

Hamid: It depends on your perspective. If you look at the semiconductor manufacturing floor, the fab or foundry operators are looking for overall throughput and yield, and what are the data points you need to optimize that. For a particular tool manufacturer, you’re looking to extract the data for machine tuning, machine bring-up, and just the analytics on that. People are looking for design improvements. For all of these different use cases, it depends on which AI tool you bring, how much data of which type you need, and what are the data security and privacy concerns. They all matter. Unfortunately, all of this gets thrown into one thing, and then the lowest common denominator holds. And hey, there’s customer data exposed in the fab so nothing can come out. That, unfortunately, slows down progress. But if we divide these into different pieces, we can start making progress across some of the aspects of the manufacturing floor.

Gupta: Generating lots of data is good. But having meaningful data, which is required by each part of the supply chain, is where there’s a big gap right now. How do I make use of it? A packaging guy doesn’t need to know some of the specifics of the transistor models in there. He probably wants to know the thermal criteria. So you need to separate it out, and then the security automatically gets taken care of. Then the fab guy is not sick with worry that his transistor data is going to the packaging guy.

SE: And that leads to another problem, because you have all this incongruous data. You’ve got thermal data, security data, data about variation and all sorts of issues that you typically run into in the fab. How does all that data get put together into something meaningful?

Narayanan: You need a data lake or platform so all that data is there. And security is a concern. At each stage, only the required data that is need should be shared. In some cases, it will tell you the health and monitor the performance of the chip. And then the applications that are running over it may take the required data that is needed. And then you can do inferencing based on the required data.

Hamid: We also need to have better data massaging and triaging tools. For self-driving cars, the vast percentage of the data they bring back every night does not need to get uploaded. There are good tools at the edge that decide what data is needed. We need to have similar tools for the fab, because any modern fab is producing petabytes of data per day, and sometimes per hour. The vast majority of that is not needed, so more intelligent tools at the edge are needed. Otherwise, there is no way to store all that data. There’s just too much coming in.

SE: Do the silos that exist today need to be modified? And do they need to be dynamic rather than fixed?

Moore: The silos in the design process are pretty well understood. But when you look at the fact that designs are getting bigger and more complex, it’s not stopping at the chip level. It goes beyond the chip, and we have to be able to evaluate things in context. We need a data platform that can open up access to all of the different disciplines that go into creating a product. As things become more complicated, and as we want to be more comprehensive, we’re going to keep adding more things that need to get checked out. We haven’t talked about standards here. When you’re creating products, like for the automotive industry, they have a certain set of standards. Communications has another set of standards. All of these need to be comprehended. We haven’t talked about sustainability here, but that’s a new thing people care about with data centers. How are we treating the planet as our customers are designing these products? They’re asking what we’re doing about emissions, and are we meeting all of our scope requirements.

Gupta: A lot of the systems companies became chip designers. That shows how important it is to get vertically integrated. You’re building chips for your own data centers. They know the use cases, they can tune in, they can process the data end-to-end because they are the actual consumers. The same thing happened with the car industry. The electric guys pushed for it. Now, many have jumped on the bandwagon to develop their own chips because they understand what exactly they need. That’s pushing more custom silicon rather than general-purpose chips.

View part one of this discussion: Using Data More Effectively In Chip Manufacturing.

Related Reading
Smart Manufacturing Makes Gains In Chip Industry
Lights-out manufacturing is increasingly sought and used by fabs and OSATs.